Good morning, everybody, and welcome to AMD's 2012 Financial Analyst Day. It's great to see such a full room here in California with us, and we'd also like to welcome everybody on the webcast this morning. We have a number of new faces who are going to be on stage today, walking you through AMD's vision and strategy in the long term and also in the near term. We have a lot of exciting things to share with you this morning. In addition, I'd like to introduce you to Rory Read, our CEO. In addition, we have Mark Papermaster, who has recently joined the company just a couple of months, and Lisa Su, who's just been here all of 30 days, so we can't wait to hear what she's going to share with us today. Then Thomas, our veteran of the executive management team on stage this morning.
We also have two board members with us here today, Craig Conway and Paulett Eberhart, so a very warm welcome to them as well. We'll take a brief look at the agenda here, just to give you an outline of who will be presenting when. This will be followed by a question and answer session, so we'd really appreciate you holding your questions until that session, and we'll ensure that ample time is provided for that session. We will host a series of breakout deep dives this afternoon in this room after the lunch on topics that we think will be very engaging and interesting. We'll have a series of very dynamic product demos. We'll have a conversation about our supply chain, and last but not least, we'll talk about heterogeneous system architecture, something we're all very excited and animated about.
I encourage you to take note of those times and to make sure you don't miss those sessions. We have a phenomenal demo area here at this event today, but before I get into those details, just in terms of a quick preview, Rory will outline at a very high level, and then the rest of the presenters will deep dive into the inflection point that we have in front of us and the opportunity that we believe AMD will capture as part of that. We'll also go into details regarding our execution ability and the importance of flexibility and agility as part of that. Then also our presenters will go into a series of other items for us this morning. The demo area that I mentioned is very dynamic.
It's down the hallway here for the folks in the room, so if you haven't been there yet, I definitely encourage you to do so over the lunch break. There's a lot of people there to help you walk through our product portfolio. In fact, we've reenacted a home environment to show you very practical usage scenarios in terms of the user experience that the brilliance of AMD products provides for you. As you take time here to carefully review our forward-looking statements regarding this event, just a few housekeeping items. We'd appreciate it if you just stick your phones on silent mode, please. As I mentioned, keep the questions to the question and answer session, and we'll address all your needs and requirements at that stage. We will have a lunch followed by the breakout sessions in this auditorium.
Last but not least, before we kick off the day, this is my Oscar moment, and I promise I'll be brief. I want to thank all the people who've put so much tremendous effort into making today the success that it will be. The investor relations team are the backbone of this event through Diana, Elena, and Armina, so thanks to them. They're also supported by a huge environment in the family that is AMD. Many thanks to our PR folks, our legal teams, our marketing teams, our facilities teams. The list goes on and on and on, but suffice it to say thank you, and I'd like you all to give them a round of applause as we now head into a video followed by Rory Read.
It's our time to step out of the shadows, to move forward. We're willing to take on the challenge. The reason people come to AMD is because they have a passion to innovate. There's a huge disruptive event that's emerging. It's about low power, graphics on the cloud, disruptive server play. Ecosystems are beginning to merge in different areas, and we know that the cloud will continue to drive a trend that's just beginning. Commitments matter. It builds trust. We do what we say, and we own what we do. It's our opportunity to reach out and seize this moment. It's our time to act right here, right now. It's our time.
Welcome, and I'd like to say thank you for joining us today at the Analyst Day.
It's an exciting time for me personally after five months to be up here and share some of the thoughts that we want to take you through during the day. I also heard that we have a room full of analysts, and I know there's a big game on Sunday. For a show of hands, I need a little interaction here from the crowd. Who's winning the game? All right, page. No, it's a show of hands, not a voice, Leslie. All right, show of hands. Who says the Patriots and the Niners are not playing for those in California? I know you're depressed. Show of hands, Patriots. OK, give me the Giants. Ooh, it's pretty close, but I have to say the analysts slightly picked the Giants. My wife's going to be depressed. I'm delighted to be here. There's no doubt about it.l.
In the first five months, we've made a lot of progress. We see an important inflection point occurring in the industry. It's one of the main reasons that I chose to come to AMD. Our company has a long history and track record of innovation. It has some of the best and brightest minds on the planet. As I came here five months ago, I can tell you I'm more excited about the opportunity today than when I started five months ago. That inflection point is going to be driven around consumerization, the cloud, and of course, convergence. These three Cs are occurring at a very important time in our industry, one where we'll see the breakdown of proprietary control points, those control points that have dominated our industry for years and years. When that occurs, shift happens. Shift is good.
Share change will occur, and the status quo will break down. I think AMD is uniquely positioned to tackle this opportunity, this inflection point, to embrace it and to drive forward and take advantage of it. Let's look at the next slide. What are we going to talk about today? I'll talk about how we've been getting AMD fit to fight over the past five months. We've got to execute. Clearly, to create value and to differentiate ourselves in the marketplace, we must execute better. It's been a key focus. We've made progress, but we have more work to do. We've got to first execute with excellence. As we do that, we build trust with our partners, our investors, our customers. Trust is the bedrock of growth. That's the future for us. That's the first lift of value we can create through better execution.
That's what we're doing right now to reposition AMD. That's why we took the actions around the organization, the leadership, how we improved our cost position, how we're structuring the company. It's all around getting this company ready to compete, and to compete in a big way as this inflection point occurs, because that's the second lift of value creation and differentiation we can create. We have to skate to where the puck is going. We have to go capture that inflection point and embrace it. That's the opportunity in front of us. We'll talk about those changes in the industry, those three Cs.
Mark Papermaster, clearly an industry veteran with a long track record of knowledge and background, he'll take us through what we're doing to create an ambidextrous, agile architecture that will open up the business for us to create those solutions that make a difference for the customer. It's about the experience that we create, how the world has been done, and how processors and how this industry has moved in the past 25 years is changing. It's changing because of the marketplace. Mark will begin to talk about the architecture. Then Lisa will come on and talk about a focus around the customer, creating a market-driven company, not one that's in an unhealthy duopoly focused on the incumbent or on the most complex technology they can capture, but instead focused on where the market is going. How do we create value?
You know, there's no mistake that Brazos is our most successful platform ever. It's a 73 mm chip, 30 million of those APUs sold. They're easy to manufacture. Boom, boom, boom, boom. Customers love them because of the experience, the graphics, the kind of work they're going to get from the cloud. Lisa will talk about our roadmap and how to capture that. Of course, Thomas, our veteran, veteran only two years, two and a half years, he's our veteran. He's going to finish what he started here. We've been repositioning the balance sheet, getting fit to fight, creating a balance sheet that we can work from, delivering on our commitments to you and to our customers. That's key. He'll take you through that.
What we'll do throughout the day, I hope, is to share with you the ideas, the approach, and how we see the opportunity to capture this next wave of success. Let's move to the next slide. There's no doubt that as we look forward and we get fit to fight, I believe in my core that you need to improve execution. It's the first lift of value creation we can do in the company. We execute better, people can lean into the wind, they can make bigger bets on AMD, and they can go after it more. How do we create that kind of execution? You build it at the foundation. The foundation, you create a culture. We call it the AMD way, one focused on ownership and commitment. I do what I say, and I own what I do. I make a commitment, I deliver on it.
That culture has also got to be based on the market, and how do I drive that market? How do I understand where the market is going? It's about creating value. If you create value for your customers, that's how you unlock the value of your company. That's why we're here. The customer, that's the only reason we're here. At the end of the day, it's about innovation. It's about leadership. I don't want to hear about being a fast follower or tracking this person or that track. It's about stepping out of the shadows and leading. These inflection points that we're seeing are our opportunity to grasp that movement. That inflection point creates the next wave of innovation. That wave is already beginning to emerge. We have to embrace it and capture it. You create that culture, then on top of that, you lay it. That's the foundation.
You lay in a strategy, a strategy straightforward, easy to understand. It aligns every person, all of what, 10,000 plus AMDers across the planet. That strategy needs to align us and drive us in that singular direction. That will drive us to get better execution across the portfolio. We'll finish what we've started. We've made great progress in the balance sheet at deleveraging our position, putting ourselves in a position where we can look at our portfolio. Where's the IP that we need in order to capture this inflection point? How do we add it? Where do we want to make our bets? We took action in terms of our cost structure back in November. That was in place to get ourselves fit to fight, but also to go after the investment areas we need. Next chart. Let's talk about what's happening in the industry.
There's no doubt, and we see it every day, from consumerization, this idea that devices are going to blur between segments. Clearly, the next 1 billion customers, what, there's 4.5 billion people on the planet not connected today without this technology. The next 1 billion customers, they're going to be in entry and mainstream price points for sure. They're going to be in the emerging high-growth markets. Consumerization is going to drive a whole set of new products. The cloud, the cloud is going to be how we deliver data and information to these different devices. It's going to fundamentally change the way data centers are constructed. I was CIO in IBM and the PC Group and in the Global Services Organization. I have experience from that perspective of what old data centers look like.
Today, they'll be tailored to huge amounts of data, high-scale volume transaction, and the processing and the kinds of power envelopes that are needed will be significantly different. That dense kind of cloud solution is going to be fundamental. You think about convergence. Every device, every screen is going to become interconnected the way applications, data move across those devices seamlessly. That's our opportunity. How do we link that material across those devices? Can we create an architecture, an SoC kind of strategy that allows from a 50 mm chip to a 300 mm chip, the ability to link and deliver that time-to-market, that kind of graphics experience, that low-power solution, and allows our customers to link across those devices? Those are the opportunities in front of us.
We see the new ecosystems emerging, the days of that proprietary control point dominating 95% or 98% of the volume of a particular set of clients. That's beginning to change. That day is beginning to end. A new wave is beginning to emerge. Who thought two years ago we'd talk about Win 8 on ARM devices? It's going to happen this year. Those devices and the way devices are used from tablets to ultra-thins, all the way through, this is going to occur, this convergence. Next chart. Let's think about this. We've seen these inflection points. This is one of the real beauties of being 50 years old. There's not many, OK, but there are a few. There is the beauty of being 50 years old, as you've seen inflection points before. When I joined IBM in 1983, you know what the dominant proprietary control point was?
It was MVS and SNA networking, system network architecture. It dominated the world. It had the gross margins, the market capital. Then this wave began to emerge. It stagnates first, then a wave begins to emerge. It was client-server first, then it went into mobility, et cetera. We're seeing another inflection point. One of the reasons I came to AMD is because I believe that that inflection point was occurring now. We can debate whether it's overnight or over a three or five-year period, but it will occur. Those proprietary control points and pricing will fundamentally shift. As it does, that'll open up opportunity for us to embrace that change, deliver an architecture and set of solutions that plays to our hand that creates value for the customer. We're going to be able to fly under a pricing envelope that's created by the old model. They'll protect the status quo.
We want to embrace that change, whether it's in the server environment or the client space. This convergence, it's happened before. It's happening again. It's why I'm here. Let's take a look at the next slide. If you think about it in terms of inflection points, this is what I'm talking about. In terms of 1983, when I joined IBM, I remember they took us into a room, and they said, you know, by 1990, we're going to be a $100 billion company. At the time, it seemed very possible. Things changed. The inflection point changed. Market capital rode it. Now, great companies, they bounce back. They jump the tracks, and they find the next wave to go on. They've gone on to be an even better and fabulous company. As that inflection point occurred, new entrants emerged and began to catch that client-server wave.
That inflection point redefined winners and losers. Shift happened. As it did, market share changed. That was the opportunity. Now let's look at today and the next slide. Today, we're seeing the same thing. We begin to see, in the marketplace, some of the market leaders struggle in terms of their market capital, their flows, et cetera. They begin to flatten out in terms of the result. A new set of players begin to emerge in this. It's early in the cycle. There's no doubt about it. We've got to be early in terms of seeing where the opportunity is and begin to change it. Clearly, as I said in the beginning, I don't want you to get the wrong focus here. In 2012 and even into 2013, it's about execution. We need to deliver and build the relationships with the customers.
That's the first lift of value we can create. Because silicon takes so long to deliver, we need to predict and see where that inflection point is going and in parallel begin to change the architectures and the business models to capture where the puck is going. It's two pieces. One lift of the first piece of value around execution in the current business, and the second lift about where the market is going. This inflection point has already begun. People said that the mainframe was dead in 1985, 1986, 1990. It really wasn't. The proprietary control point fundamentally eroded, and it took a period of time. We're going to see that same erosion here today. Next chart.
Let's think about it. There are people that suggest that AMD is caught in the middle between this large incumbent and this emerging set of players that are coming up from the bottom in terms of that ARM ecosystem. It's quite the opposite. Think about how we're uniquely positioned with over 7,000 engineers across the planet. The experience and knowledge of the processor, graphics space, our solutions, and our IP is uniquely positioned to take advantage of these three Cs. We play in the entry and mainstream price points. It's not hard for us to continue there and take opportunity and share in that segment. We know the cloud is going to drive graphic and the way data is viewed at the customer and the converged device level. We know that with that IP today, what, 50% of the traffic on the internet is graphics or video based.
Within two, two and a half, three years max, it'll be 80%, 85%. With the cloud driving it, it's only inevitable. If it's going to occur in the emerging markets, it's going to be in the entry and mainstream price points. That's exactly where we play. If we capture it with our IP around low power and around graphics, and you know the bandwidth in some of those regions is lighter, we can create the experience that's quite different. This is our opportunity to step out of those shadows and take advantage of this inflection point and lead. I don't see us in the middle. I see us as an opportunity to embrace this change. The same thing in the server space. We've got momentum after, what, the past two quarters. We've started to see a turn in the server business in the traditional space.
We'll continue to build on that. At the same time, we'll make the bets right now in terms of how that data center is changing. With Lisa's team, they are meeting every day with customers across the planet. I've probably met with, I don't know, 250, 300 major customers and partners and channel members over the past five months. They say the same thing. They see the same opportunity. If we can create the execution and the trust level, and then we can create the architecture and the technical solutions, those are our two lifts in terms of creating value. Next chart. If we think about our focus in the past and where we've traditionally been focused, we've been focused on the PC and the server space, clearly a base around discrete graphics.
We dabbled in the embedded, and it's been primarily focused on the mature market, North America, Western Europe. Our focus has been on processors, chips. It's all about pricing. Of the four Ps, we love to pull the price lever. We've got to think about creating value around the other Ps. How do we create the promotion, the right product that wins? It's always been driven through the OEM. We've been in this model of this unhealthy duopoly almost pulled along by the gravitational force. Our team focused internally on that ultra-complex technical solution and how to capture that space. Remember I mentioned Brazos? 30 million, most successful APU platform we've ever launched. That solution has standard IP blocks. It's at 40 nanometers, a 10 minus 1 technology. It's a 73 mm chip. Yet we did outstanding with it because it created the experience that the customer wanted.
We can build on that and get faster time to market. That's super important. We don't want to be on the bleeding edge of technology where we're leading in with our chin and we don't execute cleanly because that breaks down trust. I came from an OEM. I know what it means to risk your business in terms of where you want to invest and where do you want to put your bets. You've got to bet on the organization and team that can deliver. We've got great innovation and great technology. Now we need to back it up in terms of the ability to execute. Let's move to the next slide. I talked about the three Cs in terms of the opportunities in front of us and what do they mean. I want to highlight it again. Consumerization plays to our hand.
It's going to be around the value pricing, the right experience. It's going to be in the entry and mainstream price points. That next billion customers, right there. It's going to be low power, there's no doubt. I don't care if you're talking about tablets or fanless. It doesn't matter. What matters is that we create those thinner devices, the ultra-thins. It's nothing new. It's thin and light. It's going to continue. What we're going to do in that space, and Lisa will share a bit of that strategy, we are already doing reference designs. Our next generation Trinity APU has more design wins at this point of the cycle than our record-setting 2011 year with the Llano chips. That kind of focus on creating the experience, the right kind of pricing, the right kind of solution, thinness in terms of the graphic representation, that's the future.
Same thing on cloud. We can create the right kind of serving technology in the traditional server space where we play so strong today around virtualization, around the database work, around high-performance computing. If we embrace where this next wave of cloud serving is going to be in a dense type solution with an ambidextrous architecture, it should allow us to embrace that low power. That's what's holding back those data centers. It's about power. It's about space. It's about performance per watt and square footage. Convergence, clearly, here's an opportunity if we create an SoC architecture end to end from 50 mm chips to 300 mm. We position ourselves to put those products across devices and across in adjacent embedded spaces. Don't misunderstand me. I'm not suggesting that we miss our focus. Our focus in 2012 is around execution, delivering on our commitments every day, quarter in and quarter out.
Build that foundation, that bedrock of trust. That's the key to unlocking that growth. At the same time, invest in terms of where we think the business is going. We'll begin to show where we're taking the business in terms of that inflection point. We'll show it over time as we execute and deliver on that. Next chart. Where are we taking the business? What's next for our company? There should be a next chart. It really talks about, I know where it is. It's got the skater shooting the puck. I love that chart. Anyway, from a standpoint where we're going to focus, we're going to double down on the climate and the mobility space. The APU is the right architecture, the right strategy. It creates the right experience at the right price point and the right value.
I'm not suggesting to dive into smartphones in a heavy, crowded space, a low margin with an execution engine that's not proper. I want to focus on client mobility, thin and light, building on the APUs, our low power, taking the power envelope down every quarter, every generation. Go tack the server space. We've seen two solid quarters in turning that business around. We want to continue to build on that into traditional space and skate to where the puck's going in terms of the dense cloud solution. Embedded, and we've already shown that we can play well embedded using that same architecture. We create that ambidextrous architecture, that agile, flexible SoC architecture. How hard is it for us to go to thin clients where we play great today or to game consoles or onto medical or communication?
I'm not saying stretch the supply line so far, but strategically pick those areas that are very adjacent, that are easy for us to create an opportunity that allows us to be accretive to the business. The core business has to be where the core business is. This year, the first lift is on execution, second lift around moving the ball forward. How do we create value? How do we win? How do we differentiate? It's about a solution, an SoC, bringing together our IP. Mark will go into this in detail to share how we leverage that. We must continue to drive lower and lower on power. It's not only because of thin and light. It's because of high growth in emerging markets. It's in the entry price points. Why can't we deliver ultra-thins in the $500, $600, $700?
Why can't we help OEMs redistribute the profit pool with solutions like the Trinity APU? That's an exciting opportunity. We want to build on our software and APU and around the heterogeneous architecture that we believe will be fundamental to this converged environment. It is going to be a world of integrating and ambidextrous architectures and ecosystems. We need to embrace that. At the end of the day, it's about creating an environment where we get time to market. We can deliver and execute every time. Mark will take you through how we want to deliver that. That creates the ability to be more responsive. We can uniquely create silicon for a specific purpose, and we can create deeper relationships with our customer set. Ultimately, we could even fundamentally begin to change the way processors are purchased down the road. Think about it.
Almost like a game console model, this could occur. Next chart. From a standpoint, I believe that AMD is uniquely positioned to take advantage of this inflection point. I believe it's our time. Our opportunity is here right now. We have some of the world's best engineers, some of the greatest minds on the planet thinking about the technical solutions as we integrate across these architectures. How do we capture this inflection point? We're not afraid of it. We see it as the opportunity it is. That next wave of innovation is in front of us. It's our opportunity and our time to differentiate here around cloud, the consumerization, and around convergence. Our market is here in front of us.
This is our time to step forward, marrying the needs of the market with our innovative capability to bring technology to market and to do it in an agile, very efficient model. This is our opportunity right here, right now. We need to create a company that's built to win, to build a company to last. First, getting fit to fight in 2012 and 2013, creating in parallel the architecture that opens up this opportunity. Two levels of value creation and the ability to compete effectively in the marketplace that is fundamentally changing. That's our opportunity. That's the new market, and that's the new AMD. Thank you for your time. I got so fired up, I got to introduce Mark Papermaster next.
Rory, thanks very much. It's a pleasure to be here today, and it's an absolute thrill to be at AMD. We have phenomenal talent at AMD.
There's phenomenal IP at AMD. I've had the pleasure over the last three months of getting to know the team, understanding the depth of that talent, understanding where the IP is, and really listening to our team as to what we can do to retool to deliver that value more effectively to our customers. It's an exciting story, and that's what I'm going to take you through here today. Let's jump in. That IP that I was referring to is phenomenal compute graphics IP that's been developed over years. It's been developed through AMD. It's developed through acquisition over time. It's rooted in all of our products. When you look at what we've done, we've been delivering that IP year- after- year with strengthened relationships with our customers, deep delivery with our ODMs and our OEMs, combining the platform capability, combining the software enablement with that core IP.
That's a deep relationship that we've been building that we need to leverage. What I'm going to talk to you today about is how we deliver this more quickly. Rory said it earlier, how we're going to bring more agility to the way in which we operate as a development team at AMD. To do that, you actually do have to architect differently. You have to take that IP. You have to take those crown jewels, and you architect it so it can be pieced together in an agile SoC methodology. That's what we're doing. We'll show you what the result is of that and how that allows us to nail our customers' requirements more quickly and more effectively to the workloads that they care about. Let's talk about what we mean here in terms of the trends that we said earlier. Consumerization, the cloud, and this whole convergence.
How does that play into that change that I talked about, that change in delivery and IP? Frankly, it's the driving force. It's that inflection point that says we have to change the way that we deliver technology. Why? Because the experience is different of our customers. That device that you have in your pocket, your client device, your smart TV that you're using, you know, all of these, our expectation, whether we're at home, whether we're at work, we want a consistent experience. We want a consistent compute capability. Frankly, we want to persistently get at that data wherever we are, whatever device we're on, which means we're tying into the cloud in a connected way. We need a consistent way to get that in a highly effective manner, very quickly. That impacts the way in which you deliver technology across those segments.
You have to deliver that technology in a quick way, in a tailored way to each of those segments. That's what we're doing. That's why we're focusing on platform solutions. It is the end customer. Those markets that we have to be able to deliver that technology with competitive differentiation. At the end of the day, it is that IP war chest and how you enable that to deliver the platform that makes the difference. In doing so, we'll leverage that partnership that we've had with our ODMs and OEMs. We're going to continue to build on that, but we're going to do it in a far more ambidextrous, flexible way to deliver the solutions. That's what's different about what we're doing here. It turns out that you change the execution model to deliver that flexibility, to deliver that speed.
That's the engineering that we put around, how you architect each of these IP blocks, how you deliver the time-to-market, how you bring more quality as you change that methodology into, basically, a best-of-breed SoC methodology. We're bringing that right on top of the differentiated IP that we've been building on for years at AMD. Let's talk about that. What I love is the first example of unlocking the power of the IP that we've been working on. It's something we're really excited about here at AMD. It's called Heterogeneous System Architecture. When you look at how we unlock this value of the CPU and the GPU IP that we have here, it's about getting it to work seamlessly, making it easier to take advantage of these compute capabilities.
It turns out that when you can allow these engines, the CPU engine and the GPU engines, to work together more effectively, you actually get tremendous speedups in terms of how you solve problems. Actually, you can exceed Moore's law in the rate at which you're providing this type of improvement. We've got examples of unzipping files, face recognition, search algorithms that you're running. When you bring these types of day-to-day needs that we all have, whether it's running on the cloud, whether it's running on your client device, and you apply this combination of CPU technology working seamlessly with the GPU, it's phenomenal. You say, Mark, that's not new. There have been these types of engines. They've been around a while. What's different about what you're talking about? What's different about AMD's approach here?
It's fundamentally different because the reason that you haven't seen this catch a wave and take off in the industry is it's been hard to program. It's been hard to unleash that value. What we're doing is creating a construct, an architecture that makes it very easy. It's going to make it as easy to take advantage of that CPU and GPU engine as it's been for years, writing high-level languages that you see running all the applications that you have day to day on your smartphone, on your tablet, on your client device. That's what this does. You know, it gives you massive speedup. We're seeing on, let's say, an OpenCL application, 125% speedup. We're seeing applications that run at half the power with HSA, with these approaches, this architecture applied versus conventional techniques. That's the beauty. Lower power, more performance at lower power.
It's really going to unlock developers to take advantage of the full compute capability of APUs. The way that we're going to go about this is in an open way, open standards. We have been huge supporters of open standards at AMD. You're going to see this in the Heterogeneous System Architecture approach. We found that the way to really bring momentum around a new technique like this is to bring the industry with us. In fact, our competitors have said, yes, we're open. They open a narrow segment of a piece of code, release that out there. That's very, very different than an open standard like Linux, where it's open to the full community. Every aspect of the tool chain is open to the development community. That's what we're doing with HSA.
It's an approach that will allow hardware developers, ISVs, the OS developers to come together in a way that there's a set of constructs, enablement to allow us to unleash this value. We're creating a consortium. We've walked out in the industry and talked to our developers, talked to OS, we've talked to competitors. We shared with them this proposal, this architecture. We've gotten an overwhelming response. It's very positive. We're in the midst of this. We're getting very positive response. We roll it out. We're releasing specifications. We have a summit coming up in June, which we're getting tremendous excitement around. This is what's going to unlock this value, making it easy to take advantage of CPU and GPU. What does that do for us at AMD? Mark, that's a tie that raises all the boats. Exactly. That creates the ecosystem. That's what we want to do.
We want this ecosystem. It lets our IP stand on its value and competitiveness. This is where we've invested. This is where we have competitive differentiation. We'll compete on that value with an open ecosystem that unlocks its capability to the developers, the software developers. Very powerful. This is a stepwise progression. We're on this journey. It's the APUs that Rory talked about that we've been shipping very successfully because that started the journey by bringing that CPU and the GPU onto the same silicon. It's got a physical memory, which is shared, that gets you an immediate speedup as you're running applications that could take advantage of both. That's just the first step. The next step, as we're rolling out this year, is allowing a much better management of the power as workload shifts. It might be on the CPU.
Then aspects of the application are much more parallel and can be taken advantage of in these parallel GPU compute engines, shifting back and forth, managing the power in 2012, and allowing the GPU to access the x86 memory space, physical memory space, right? In 2013, we actually go to a virtual memory, 64-bit addressing with virtual memory shared between the GPU and CPU. In 2014, we have full complement of support for this Heterogeneous System Architecture where we have immediate context switching between the GPU and CPU and really have unlocked in a very seamless way with all the software constructs underneath this how to get the full value and full savings of this combination of technologies, two different ways of doing computing on a seamless basis. When you look at why this is critical for us, this is our crown jewel.
This is the graphics performance that we've had. Year after year, we have been delivering graphics leadership to the market. Our AMD Radeon HD 7970 that we just announced and began shipping at the end of last year again provided leadership capability. Its leadership performance is four teraflops, four teraflops of compute capability. More than that, it's done at a significantly lower power. We're improving performance per watt at each turn of the crank as we deliver our product. We deliver leading-edge standards at each release of our graphics IP technology. DisplayPort, Eyefinity, the support for multiple screens, the latest in DirectX. We support DirectX 11. We push this out in a leading-edge methodology and leading-edge technology. We've got demos, which I'm hoping that you'll have a chance to see over the course of the breaks and during the day here today to see this technology in action.
We bring it out with our discrete GPUs, and we roll it right into our SoC methodology into our APUs. Let's talk a little bit about low power. I told you that that was a fundamental element in terms of our strategy because we're rolling out low power, leading-edge low power across each and every aspect of our IP and our SoC design. This journey is one we've already been on. We've been on this for several years. This started with AMD being a first in bringing key elements on board the silicon to drive drastic reduction in power and improvements in performance. We brought the North Bridge on board. We brought the GPU with our first APUs and South Bridge integration. The first step was integration of function to bring down power and improve the efficiency.
Next, we went to power management, onboard power management, which we've been shipping as well since 2010. We've gone to power gating. Turn on the power where you need it, where you're computing at that moment of time. If there's not operations going on on that piece of the design and the architecture, shut it down. Don't have it burn any power. Improve that even further by having that be very fine-grained. You break it down to the precision of where you need to burn the energy at that moment of time and have the rest shut down. What we're rolling out this year in 2012 is more advanced power management. I talked about it a little bit in the graphics and in terms of how we shift the energy from a CPU to a GPU based on where those operations are going on. Advanced thermal management.
Ultimately, as we complete our rollout of the Heterogeneous System Architecture, this provides a very fine-grained control of where the computing, it's optimization of taking algorithms, taking the workloads that our customers are running and ensuring that it's optimized for efficiency and where the power is being burned. End result, it's not about the cool technology or the features that we implement to enable these power savings. What matters is for the customer. What matters is battery life. If you look at the APUs that we've been shipping this year, it represents a doubling in the battery life of what we had just two years prior. That's what Brazos is delivering. It's a phenomenal, it's an end-user experience that changes. That's technology put into application. That's what it's about. It makes a huge difference in the data center where we're thermally constrained and you're power constrained.
It's power delivery into the square footage you have in the data center. This makes a fundamental difference to our customer, taking technology and innovation and applying it to make a customer difference and being more agile in the way that we deliver that across our entire product line. Let's go further when I talk about that kind of IP. When I say we're applying it across all this IP, this is our treasure trove at AMD. We've been working on this for years. We have a deep set of IP across delivering that experience, the multimedia IP to deliver that graphics experience in our discrete graphics, our APUs across each of the segments I described, bringing that power management across the board, bringing that into the application.
When you're running and you're using our technology for displays, right through the Codecs, right through that whole translation of the data into that visual image, bringing all of these features to bear across our core, right into our memory feature, how you're managing the memory. It's about taking the techniques that we have with low power, with Heterogeneous System Architecture, bringing flexibility and agility, defining each of these IP blocks such that they can be put together in a very agile, flexible way that's going to deliver a different experience to the customer and a different mode of delivery for the AMD company. You know what? We're going to give you an example of how we think about that. It's really, I call it a horizontal and a vertical mode of delivery of that IP and the delivery of the technology and strategy that run. First is horizontal.
That's about piecing together that IP I showed you. How do we bring that together? What we're doing with the SoC methodology is we specify each of those IPs with a consistent standard way in which it can be put together with a consistent methodology, one methodology across the AMD company as to how that comes together, giving us that flexibility to mix and match all of that AMD IP I just showed you in a very faster way. That's a fundamental change. It's going to make us more agile. That's the horizontal that we gain, the horizontal efficiency that we gain in putting these solutions together. It's actually more than that because you've seen what we're already doing with the CPU and GPU. With that flexibility, we can tailor solutions in a much more agile way to the demands of our customers.
What that means is as we've architected this new SoC methodology with these consistent interfaces, it also plays and supports standards that are out there for ecosystems that exist outside of our world, outside of the internal IP that we've been developing at AMD. That allows us to very easily pull in third-party IP. We don't want AMD engineers working on off-the-shelf industry available IP. If our customers need that, we're going to be able to pull in that third-party IP directly into the SoC methodology that we're putting in place. It's a straightforward concept. It's very powerful. This is what we're executing. Let's go to the left side. What do I talk about vertical with this vertical integration? It's the strength that we've been working on for years, developing those relationships with our customers because they don't just want a piece of silicon from us.
That's not what the AMD company is about. We deliver solutions. We deliver the Tor technology. With the SOC, we'll be able to target those engines to our customer needs in a more faster way. It's about the software enablement that we bring with us. We have a massive investment in the software and the enablement around each of our IP blocks that we develop. It's about the platform. We build reference boards. We test this out. We deliver functioning platform solutions to our ODM and our OEM customers. We can do so in a specialized way as we're doing with game consoles and services to help them test and give our customers more efficiency in delivering those platform solutions and the software enablement that they need. It's about the horizontal, the flexibility, and piecing together our IP as well as external IP. It's about the vertical.
Nail the solutions that our customers need and they're asking us for and doing that in a flexible and agile way. What I'm going to do now is talk to you about the perfect example. It's what we're focused on to take what I just described to you and how we're changing the way at which we deliver IP, the methodologies we use, how we enable this with Heterogeneous System Architecture to an application that we're very excited about. It's a data center. The data center is changing. Rory talked to you about this inflection point that we're going through in the data center. I've seen this as well. I have many, many years working with our data center customers.
It is just very clear to us the value we can provide as we take this flexibility, this is what we call ambidexterity, to be able to mix and match different solutions of our IP. You look at what's going on in the data centers. For many of these cloud workloads, you have huge farms. They're running tailored workloads. They're running web server. They're servicing search engines. They're running more specialized. Some are running high-performance compute applications.
We need the ability, and we have it with the approach we're taking, to provide flexible, tailored compute solutions on those types of applications, but do it in such a way that we're leveraging the investment we've made not only in the IP, but the low power so that we can take our designs, optimize for low power, and add IP that allows them to interconnect in a very dense fashion, very efficiently, right? Integrate this together. It is that aggregation when you put that compute capability, tie it in, working with others, bringing the I/O, bringing the storage capability, bringing the solution, the platform together with our IP and the flexible methodology we've developed. We can provide the solutions over the target that the data center customers need, given the inflection point, given the challenges they're facing. We understand those workloads. We've been working with them for years.
We're tailoring our solutions so that we can deliver our IP, our low power, our acceleration to the data center. It is going to be disruptive. Data centers, mega data centers in particular, the cost of running them is measured in power. It is measured in the computing per square foot. It is about lowering that total cost of ownership. That is what we're going to do. If you look at how we go do this, it is simply about leveraging the capabilities which we are developing. It is about executing, bringing it together on the plays centered over the data center in terms of what they need. As we take that flexibility and that IP blocks and the Heterogeneous System Architecture, we bring that together in a set of solutions, but interconnect it in a very efficient way so you can create hundreds, thousands of these compute elements together.
This is what's fundamentally different. You bring efficiency. You bring efficiency to computing. It is going to be a differentiation. The feedback we have is this is the right strategy for AMD. It is leveraging our IP. It's leveraging it to solve problems that our customers care about. This is the track run. We're bringing this technology, focusing on the data center. It's the immediate focus, that change in strategy I described to you, to hit the market inflection point that we see in front of us. When you look at how we do this, how do we tailor for custom workloads? It's pretty straightforward. It's about architecting. It's about taking, you saw all of that IP I showed you on the previous slide that we've been working on developing for years. It's our crown jewels of the differentiation. It's architecting it right up front so it can be flexible.
You take the x86 IP that we have such a rich heritage, 64-bit computing for many years. We've been optimizing this. We have cores which are tailored for high performance per watt. We have cores that are tailored for the low power applications. We have two lines of cores. We're architecting them such that we take all the x86 aspects and we bring it right into that core IP block so that x86 ISA can just be matched up with all the rest of the IP that we have. Flexibility as we drop that x86 and mix and match it with the rest of our IP. That approach of changing how we architect and how we interconnect our IP opens up the ability to bring in third-party IP. If our customers had some unique IP that they need, we're going to work with them. We can create purpose-built silicon.
It's about leveraging our offerings, our IP, understanding the needs of our customers, and leveraging it. Leveraging that and leveraging the ecosystem that's out there that's creating IP, complementary IP to what we have here at AMD and bringing it together to solve the customer workloads. It's compelling. It's absolutely compelling. The feedback we have, I just spent two weeks between CES and travel I've done sharing this story with our customers. It's a very positive feedback. This is what they want to be more efficient to solve the problems and, frankly, to differentiate themselves. It's a very compelling story. You've got to execute differently. We will be executing differently. We've already embarked on this journey, aspects of it like low power and our march to Heterogeneous System Architecture to further unlock our CPU and GPU capabilities. We've been on that march. We're doing even more.
We're changing our SoC methodology to be rigorously modular to enable us with a best practice, consistent methodology across each and every one of our design teams to be able to put this together in a much more agile way and to have true IP reuse. We're architecting our IP for the target segments that we know that it will go. We've architected up front the interfaces it needs to be able to drop in each of those applications. Secondly, it's about how you bring those solutions together in a way that's more quick in time to market. We're providing a more parallel approach to our hardware and software development. This is also a fundamental change. It's really enabled by the advent of highly capable emulation systems that are out there in the market today.
What we're doing is defining the features for our software for that enablement around the platform. We pulled all this in concurrent to when we designed that SoC capability up front. What does this mean? When you line up all that definition up front and pull that together, pull that development together in parallel, it actually is like a set of dominoes that fall to additively increase your time to market, additively increase your quality. Why is that? When you line that up in parallel and you do this in a more cohesive way, you get the requirements in working cohesively, validating in parallel while that long lead time silicon is in development. You're in parallel honing how that runs in a customer application, ensuring you hit the performance goals, validating the features that you need right out of the get-go when that comes back.
You end up with more parallel development. It's just a pure shrink of that cycle. When you go to deliver this to the customer, that validation is shorter. You've already run those features. What this means is we're more rapidly going to get our IP in the hands of our customer. Thirdly, I talked about low power. As we drive low power, it's all the techniques I talked about we're putting into our silicon. It's about how we manage across the platform and the stack and how we optimize performance per watt across each of our IP, across the SoC as we integrate it, and how we manage that power.
It's the benefit we have of having been in the platform business for many years to not think as a pure silicon provider, but to think as a platform provider that we are to end up with the optimized solutions for our customers. This is a change. We're leveraging what we have done. We're leveraging the experience we have, the deep talent that we have, but we're becoming much more agile. We've had that talent, and we've created that highest performing, historically, that highest performing capability in the markets that we've served. We've gone after that last 2%- 3% of performance. It's been a longer, historically, it's been a longer development cycle. It's been a several-year development cycle the way that we had been operating. We've used the absolute latest technology node with a new architecture that we've been developing.
Frankly, we've had some challenges as we stacked all of that together. We've had a couple of areas where we had, therefore, challenges when we stacked risk upon risk. What our customers need, as evidenced by the Brazos example that you heard from Rory, is they want agility. They want that SoC solution. They want it tailored to the workloads, the problem set that they're solving. They want it more quickly. That's what we're doing. We're simply listening to our customers, changing the way that we take that wonderful talent we've had, the wonderful IP that we've developed, but bring it to our customers in a fundamentally different way. It's more quick. It's more agile. It's tailored to the problems that they want solved. I talked to you about the play on the data center. That's where we're starting.
We're rolling this out across everything we do, the methodology, the approach. You'll see it implemented across all of our segments. At the end of the day, in addition to speeding us and getting our IP to market, it makes for a different AMD in terms of the ability to partner with us. It's just going to be easier, right? It's easier for us to understand those needs and rapidly take our IP, bring it together to solve those customer needs. It's an exciting time at AMD. I am thrilled to be here. We are running with this new play. The team's very excited about it. Look forward to you hearing more detail over the course of the day. Thank you very much for your time. Appreciate it.
Our program continues here. Once again, Ruth Cotter.
Grace, welcome back, everyone, as you start making your way back into the room. Welcome to everybody re-engaging on the webcast. We're going to continue our program here this morning. I'm delighted to welcome Lisa Su to give us a very solid overview on our product portfolio and the direction that's going to take. She needs absolutely no introduction, as I know many of you in this room are very familiar with her tenure. Lisa.
OK, good morning. It's great to be here today on behalf of AMD. As you've heard this morning, I'm the new kid on the block. It's really an opportunity for me to talk a lot about where we are going with our products. Rory talked about our vision for the company. Mark talked about the technology direction. I'm really going to focus on our products and how we take our products to really grow market share in this very exciting market that we're in today. It's also been an opportunity for me to take a step back and look at how I stitch together this great product portfolio that we have into something that's really, truly differentiated in the market. Hopefully, I can give you some of that as we walk through the presentation. I'll start first with where we think the market is going.
We already talked about these inflection points. It really is a very special place in the semiconductor industry today. Next, I'll talk about our technology leadership and then spend the bulk of the time talking about our product roadmap. I know that many of you have been waiting for an update on our product roadmap since it's been a while since the last public update. We'll go through a lot of details of where we have made some changes and where we believe we're going to win going forward in the next couple of years. I'll talk about what the future holds and where we're going to go in terms of the new markets that we're pursuing. Let's start first with the market trends. This is an extremely exciting time to be in the semiconductor industry.
We've all been here for many, many years, and we've seen how we've gone from the PC era to the mobile computing era to the cloud and the internet devices. Where we are today is really the era of convergence. What that means to me is imagine having your data anytime, anywhere, any place, and in any format, all interconnected with devices that really have that convergence. That means tens of billions of devices are going to be sold every year. That's both on the consumer side and on the infrastructure side, and that creates great opportunity for us going forward. I know a lot of people have talked about, hey, you know the PC market is not such an interesting market anymore. It's being taken over by a lot of different areas. That might be true if you're talking about just the very traditional piece of the desktop market.
When you look at the segments and where the segment growths are, there are so many places where we can grow in this industry. If you look at Notebooks, we'll talk about Notebooks a lot and the ultra-thin form factors and the light and slim versions and what we can do in the emerging countries. That's tremendous growth for us. When you talk about the server market, beyond traditional IT, all of the new workloads and the mega data center and the cloud and what's going to serve all of those internet devices that are being connected, that's growth opportunity for us. When you talk about the embedded space, lots and lots of opportunities in the embedded space. Those are spaces that are very adjacent to the technology that we have today, so that's room for growth for us. Then you talk about tablets.
We know that tablets, if you look from a revenue standpoint, being a smaller sliver, the unit volume is very, very high. That connectivity and that technology go across the markets. You see that all of these are tremendous growth opportunities using the same basic technology for compute and visualization that we're very known for. Rory also talked a little bit about this. If you look at where the next billion users are going to be, there's tremendous market opportunity in this area. When you think about the last decade really being the growth of mobile phones and smart devices being connected to the internet and really driving growth in the mature regions, when you look at the emerging countries, just think about all of the opportunity for productivity, for education, for governments, for modernization of the infrastructure. All of that is tremendous growth opportunity for us.
If you think about what really drives those areas, it's not about the nanometers or the pure bleeding-edge performance. It's really about what user experience we can really drive. When you think about that, you're going to hear us at AMD talk a lot more about the experience. What
are we trying to drive? It's not just technology for technology's sake. It's how do we take technology to something that means something to the consumer and to the marketplace. When you think about some of the user trends, you know, think about our own user patterns when you're talking about a laptop today. It has more than enough compute performance for what we need to do in our PowerPoint, in our word processing. When you talk about driving the next generation, it's really about the natural user interfaces, the touch, the sensing, everything that these new devices bring to us requires computing power in a different way. When you think about our display technologies and the idea that you want to be as real as possible, and that realism comes through in these immersive displays, that drives computing power that's different from just straight regular computing.
When you think about social gaming, that's another aspect that drives tremendous capability that's different from what is in today's computers. Now, when you go into the infrastructure, we want to connect all of these devices in the cloud, where you see data everywhere, all interconnected, being able to get access anywhere. In the collaborative environments, we want to take these video systems that are tens of thousands of dollars today and put it into your $500 PC. That is the way that we think about driving the user experience in the next wave of computing. When we take all of these market trends, and I've really tried to figure out, you know, how do we put simply what is AMD's product focus? AMD's product focus is really about aligning to what the market needs. It is about creating an exceptional user experience.
That is what's going to drive the consumer. That's what's going to drive the computing capability. It's a user experience across all device categories. What you hold in your hand is the same as what you use at home, is the same as what you use in your car, and what you use at your business environment. All of that is one single architecture and the ability to share across those areas. It's about taking leadership, compute, and visualization. The idea is compute is practically free now, if you really think about it. With the advancements of technology, you can really go very, very far. We want to bring that to lower power and lower cost points. That's what we can do with our technology. Then you heard Mark talk about flexibility.
I think the value proposition that we want to bring to OEMs is one where we can really bring flexible system on chips that go across all of the different platforms and be able to really get the time to market. This is the value proposition that we want to bring to our customers and to the consumer. Now let me talk a little bit about technology leadership. I am a technologist at heart. The fun thing about technology is really figuring out how we put it into action. When you think about technology, for a company like us, you don't have to be the best at everything, but you have to be the best at a few things. That's the differentiation that we have to get into our products, into our mindset, into the marketplace. In terms of technology leadership, we talked about graphics.
Graphics is definitely one of our crown jewels. The graphics performance leadership that we've demonstrated over the last decade has been phenomenal. We talk about performance at the very, very highest end in terms of gigaflops. We'll talk about our four gigaflops machines. We'll talk about graphics in terms of performance per watt, driving it throughout our capability. We've also talked about graphics from the standpoint of giving you a different user experience. If you go into the demo labs later on today, you'll see our Eyefinity display, which goes across six screens. You can see, you feel like it's different. It's different than just what's available in today's capability. More important than that, the graphics technology ends up being the centerpiece of the entire rest of our roadmap. Graphics is a basic building block in terms of parallel processing capability that allows us to accelerate many, many, many applications.
Desktop productivity, browsing, all of that capability. This is really the secret sauce that goes into our APU line. When you talk about APUs and application processing, I've really thought about heterogeneous computing for the last 10 years. I think heterogeneous computing is really what semiconductor folks think about as the best of both worlds. If you think about it, it was first used in supercomputers, then used in gaming machines. AMD really made a big bet, bringing it into the mainstream. The idea is that you want the right computing element to be used for the right application. In 2011, we launched the AMD APUs. It's been fantastic. You heard it from Rory. We'll say it many, many times today. First to introduce heterogeneous computing in the marketplace. It had great customer reception. The performance of the Llano APU is 3x what a typical general purpose processor would do.
That's the power of bringing the processor and the graphics capability together. When you look at the roadmap, and I'm going to talk to you about the roadmap for second generation and third generation APUs, we will take that in the mainstream to one teraflop. It's really just a couple of years away. The thing about APUs is where do we think it can go in the market? If you look at the market adoption for this technology, it's been fantastic. We've shipped over 30 million APU units to date. If you look at it just starting shipping in the fourth quarter of 2010, it really has made tremendous progress. 11 of the top 12 OEMs are shipping AMD APUs. We see that continuing to grow, the momentum continuing to grow into 2012 and 2013.
Brazos, we've talked about, very low power, very nice in emerging markets, really goes across a range of applications in both desktop and laptop. It's the fastest growing platform in our history. We see the next generation going even faster. Lots and lots of good customer reception around the APUs. We continue to believe this is a strategy that's going to differentiate us going forward. I also want to spend some time on server. If you really think about where the growth for us is going to be, it's really the client side when we look at the growth around mobility and the server side when we see growth around the data center. Server is a place that we are absolutely committed to be successful and win.
When you look at the architectural investment that we made with Bulldozer, Bulldozer was a bit ahead of its time in terms of really focusing on multi-threaded workloads for the server market. It was designed for server performance workloads. We put a lot of effort into the revolutionary power design. I'll show you what some of the customers are saying about the server architectures that we have. Once again, our focus in server is not on single-threaded performance alone. It's really on total system performance, as well as total system power. That's where we're committed in the server market. Now let me move to the product portfolio. There are a lot of details in this. I'll try to walk through each one of these to give you a sense of where we're going in the product portfolio.
Starting first with our strategy, it was very important to pull together AMD's product strategy on one piece of paper. It's not a complicated strategy. It's actually a very simple strategy. It actually comes together across each of our market segments and includes why we think we're going to win. Starting first on the graphics side, it's very, very clear that leadership graphics performance is absolutely critical to us, both in the discrete graphics market, as well as the capability to bring us through that visualization experience. This is going to be a very, very key part of our differentiation going forward. Our goal is to leverage that IP throughout our entire product portfolio. You're going to see that that graphics IP is now going to go throughout our APU line and in every market that we serve.
We believe driving this is going to give us unique differentiation and give the consumer unique value in that capability. On the client side, it's all about the user experience. It's all about taking that user experience to lower and lower powers. We want to hit the volume sweet spots of the industry. That includes mainstream entry. Yes, that's going to include tablets in a big way, because that is where we believe there will be growth in the market. When we look at the server market, I talked about the importance of the infrastructure. We believe the infrastructure is changing. We believe that new workloads are where the growth is going to be. We're going to take that leadership in performance and performance per watt, but really customize solutions, as Mark talked about, in the data center space so we get that together. Underlying all of this is execution.
Rory talked about the first wave execution. Mark talked about execution. I'm going to talk about execution. At the end of the day, this product roadmap is a roadmap that our customers can count on. It's a roadmap that we're going to build upon so that we have a very, very strong, very flexible SoC system that we can really get time to market from our customer standpoint. Now let's start with the client and graphics roadmap. As I said, there's a lot of information on these charts. I'll try to go through each box. Starting first at the top on the graphics side, we're very, very pleased with the graphics roadmap that we have today. Our Southern Islands launch started just before Christmas.
We got tremendous reviews on our Tahiti product line, and we believe that it has clear leadership in the industry with a top to bottom discrete roadmap. I'll show you a little bit more detail on the next chart. Our second generation Trinity APU, we've talked a little bit about it in the press, but I can give you more details today. We certainly believe this has tremendous, tremendous value proposition, both in terms of performance and the performance that it hits and in terms of the power envelopes and the systems that it can go into. The design wins have also been spectacular. In our low power and our ultra low power lines, there is a change to our roadmap where we are continuing with our very, very successful Brazos line, going into Brazos 2.0.
This is actually a very good value for our OEM customers because they're able to take all of the infrastructure that they put with Brazos and really take it into the next generation. You'll see that we're extending our roadmap into ultra low power. Honda will be our first tablet-based system, and we will see tablets with Windows 8 out there with AMD silicon. Those are the key elements that I'll talk about in the next couple of charts. With Southern Islands, I will say that it's really cool to have the fastest graphic cards out there. That's something that we're very, very proud of. We've gotten really tremendous reception from everyone in the market. When you look at the performance capability of Tahiti, as I said, four teraflops, 4 billion transistors, first to market with 28 nm technology, support for six displays. All of the key elements are there.
This product has been very, very successful. We're shipping it right now. You will see before the end of the quarter the announcements of our performance and our mainstream line with Pitcairn and Cape Verde. All of that gives you end-to-end leadership in discrete graphics capability. This is something that we are very, very committed to and will continue to drive hard because we believe this is a key market and a key way for us to leverage our IP. Going into Trinity, this is really our second generation APU. This utilizes our Piledriver cores. Piledriver is the next generation x86 core, up to 25% performance over Llano. It also incorporates new graphics from our Northern Islands and Southern Islands capability. We have the graphics core from Northern Islands. We have the video IP from Southern Islands. That gives us up to 50% graphics and compute uplift.
When you take a look at the numbers, you can really think about Trinity in two ways. You can either double the performance at the same power, which is great, or you could take the same performance at half the power. We can go from 35 W from Llano down to 17 W from Trinity. That gives us tremendous performance per watt leverage, which really kind of takes this class of machine into a different space. From a battery life standpoint, over 12 hours, we feel really, really pleased with where we're positioned from a power standpoint. The important thing here, as we noted earlier, is that the design wins are tracking today ahead of where Llano was at the same time. The Trinity products have started shipping already into OEMs. You'll see systems on shelves by the middle of this year.
This product is ramping very fast and will be very key for our 2012 execution. Now at CES and in the industry, there's a lot of talk about ultra thins and what we can do with ultra thins. We have a number of products that are out there in the showroom. I wanted to bring one up here on stage because this is a reference design that's built off of the second generation APU Trinity. You can see that it's using a BGA form factor for the ultra thin. It's a 17 W form factor. The thinness and the lightness of it is pretty incredible. At 18 mm , it is really one of the reference designs from one of our partners at Compal that many OEMs are looking at and will be looking at putting into the marketplace.
We believe this will bring the ultra thin form factor, which is very, very sleek and very light, into the $600- $800 price point. That's the value of the APUs. You really have the performance that you need at the power envelope that you need at the right price point. We have tremendous, tremendous belief in this platform going forward. You can see a number of platforms out there with the Trinity system. Once again, it's very, very, very, very good reception from our customer set. Now let me talk about the entry and the value lines. The Brazos 2.0 platform is a very important platform for us. We've talked about the success of Brazos and the APU. The fact that we can get platform continuity and execution and all of that for our customers was a very, very key element.
When we think about how to really upgrade this platform, we've been able to turn on our Turbo Core technology, as well as optimize for Windows 8 and add some more features to it. This gives a very nice platform continuity for something that has shipped very well in the 2011 time frame. We're also seeing a very nice improvement in battery life. That will be helpful as we go into the emerging markets. We've also been able to take a new SKU from this into the Honda platform that, as I said, will go into the ultra low power tablets. Brazos 2.0 will really fill out the lower end of our platform. We'll really, really focus on the ultra low power points there. With that, let me move on to 2013. 2012 is a year of execution. It's a year of Southern Islands. It's a year of Trinity.
It's a year of Brazos and Honda. As we move into 2013, you can see our whole lineup will be refreshed in 28 nanometer. We don't usually talk a lot about our graphics roadmap at these analysts' day. Our Southern Islands family will be moved over to Sea Islands. That will be a new graphics architecture that will increase the performance and the capability. Very, very importantly, it will also add new HSA, or Heterogeneous System Architecture, features. Mark talked about it. We are absolutely committed to making this up and down our roadmap because we believe this is one of our key differentiations and value propositions. In the APU line, we'll go to our Trinity. We'll go to Kaveri. Kaveri will be our third generation APU. It was that one teraflop point that you saw on the APU line. We'll get significant performance improvement.
It has a new x86 core in it that will get more IPC and power improvements with Steamroller. It will also have new graphics and all the new HSA features that I talked about. As we move into the low power and the ultra low power lines, here is where we really drive for more and more performance at lower power. We talked about the system on chip architecture, why we think system on chip is so important. It brings so much value. In Kabini and Temash, we'll have our first true system on chip implementations, integrating the Southbridge, the Fusion controller, onto Kabini and Temash. That gives us power improvements. That gives us performance improvements. That gives us a really system bill of materials space and cost savings.
That will be off of our Bobcat Core series, which will be upgraded with Jaguar, which will give us more performance and more lower power. It will also really give us an opportunity to scale these HSA features across our entire roadmap. When we talk about our belief in APUs and HSA really being very, very critical for our roadmap, you see it in practice in 2013. This roadmap is very much in execution and is very clear that our focus is on time-to-market in this area. Now let me talk a little bit about servers. We've talked about the opportunities for growth. We are really looking for, I think my job, as Rory defines it, is to take all this technology and grow market share. Server is a great opportunity for us. It is clear that our server market share is not all that high today.
We view that we've made a very, very strategic investment. There's a lot of opportunity for us to grow in this area. The value proposition is simple. We want to provide something that differentiates ourselves from the incumbent and really focus on value, price performance, new workloads, workloads that really take advantage of our architectural investment. When you look at what that means, these are some of the numbers in terms of where the server market is going to go over the next couple of years. The traditional IT space is very large and very important. It's not the fastest growing piece of the market. When you look at the high-performance computing workloads, the virtualization workloads, the cloud workloads, here is where we're seeing a lot of the growth. Here is where we're also seeing a lot of the opportunity. Customers are coming to us asking for solutions.
They're asking for solutions that will help them take the total cost of ownership down. They're asking for solutions that will really help broaden their ability to attack these markets in a disruptive way. This is where we've really focused our architecture. Late last year, we announced the Opteron 6200 series. This is the Interlagos platform family. Very, very significant change in our architecture. Very clear investment in Bulldozer and the architecture with the modular design. Clear that we were focused on multi-threaded workloads with the dual integer units and the shared floating point unit. The world's first 16 core x86 processor. We've gotten tremendous feedback from the market and the customers about the architecture. It is fair to say, and we all know this, any new architecture takes time for the entire ecosystem to adopt to it. We need software developers.
We need the tool vendors to really adopt to the multi-threaded capability. We're seeing progress. We're seeing real progress in the adoption rate. We've seen two straight quarters of server market share growth, albeit we have a long way to go to reach our aspirations. We really believe that this architecture is what will take us there. It's the focus on not just performance. Performance is very important. It's the focus on power efficiency and how we really put features in for the data center. Things like our power capping capability allow data centers to really decide where they want to set their overall power. These are the types of things that we think will bring differentiated value to our customer set. Some of the things that we've seen already over the first couple of months, as I said, our customers have given us very, very strong feedback on it.
One of the marquee wins was the Blue Waters win with QRAY. When you think about supercomputing, this is going to be one of the largest installations of supercomputers in the world. We'll have 50,000 Opterons in that installation. We hope as it comes together, we'll be close to one of the top few in the world in terms of supercomputing capability. That's an example of utilizing our architecture in a very differentiated fashion. We've gotten very strong performance on 2P performance for, let's call it, the bulk of the workloads. A lot of good database feedback. A lot of good virtualization feedback. These are just the beginning. As we said, it takes really the opportunity for the entire infrastructure to really adopt to the new architecture.
When you look at our roadmap going forward, our roadmap going forward is really focused on ensuring that we get smooth customer transition overall. This year, it's about Interlagos and Valencia, which will power the majority of the systems that are out there today. We will also be starting to ship Zurich later in the first quarter, which goes for the microserver and the lower-end workloads. As we move forward, we made a decision to really take the infrastructure that we have and really move it forward without change for our customers. Abu Dhabi will be a straight replacement in the same socket and the same infrastructure for Interlagos. We'll get higher performance with the Piledriver cores that will give us more IPC and more performance. It'll also be a very, very smooth transition for our customer set when we go forward.
We had in the past looked at should we add more cores or add new sockets. At the end of the day, that wasn't the right answer for our customers. What they wanted was to start with Interlagos, to move to Abu Dhabi, similarly with Valencia and Seoul. This is really where we're focused, is really taking advantage of the infrastructure investment that we've made together with our customer set to be successful in the server space. It is also really, really important for us to repeat the commitment to technology. The server technology space is one where the treadmill is very fast. You have to keep investing and keep going at a pace. The Bulldozer technology, we believe, has an extremely long life. It has a lot of features that we can add and improve on with each subsequent generation. We've talked about Piledriver in 2013.
Steamroller will also be a next generation core that will give us more performance and more IPC capability. As we go to Excavator, you're going to see some of the features that Mark talked about, about really getting a very modular, very low power design technique. We're actually starting to share design techniques between our low power core and our high performance core so that we get very, very power efficient in the overall architecture. From this standpoint, you'll see that the Opteron line gives a very long life for the investment in our architecture. These will, of course, also go into our desktop devices as they flow into the client space. It gives you an idea of where we're investing and where we believe the technology is going. Now let me spend a few minutes on the future.
We have a lot of exciting products on the roadmap: 2012, 2013, execution, execution, execution. A lot of conversations with customers. What they tell us is more products faster, please. That is certainly what we're working on. There are also a lot of opportunities for us. I really look at what we have here at AMD as a land of opportunity. We have all of this great technology, we have a lot of new markets to grow in. We just need to marry those two things together. When we look at the future directions, Mark touched on our technology investments. I want to reiterate from my perspective. Leadership IP, there are not too many people in the industry who can talk about the realm of IP that we have, whether you're talking about graphics or multimedia or processing capability or all of that capability.
It is really a unique treasure chest of IP. When you think about how that links into solutions, the Heterogeneous System Architecture really is something special. Because I'm a lifetime hardware person. When you think about hardware, you can put all the transistors you want on a piece of paper. If you can't really program it, it doesn't help. That's what the Heterogeneous System Architecture is about. Low power, a lot of people have said, AMD, do you get low power? I've heard that before I joined AMD. I've heard that since I've joined AMD. We absolutely get low power. Absolutely get it. We're going to put it into our products. Then ambidextrous architectures just open up a whole new world for us. In other words, we don't have to be religious on any particular element of technology. At the end of the day, we're a solution company.
We're trying to deliver solutions to the market. There are some architectures that are good for a lot of things. There are other architectures that are good in different ecosystems. Our job is to choose the best of each one for the marketplace. In terms of product growth vectors, I'll talk about each one of these and where I think we can grow. Very clearly, from a mobile market standpoint, we have a mission to grow in the mobile space. We've made very, very good progress with our notebook line, our APUs with Brazos as we go into Honda, as we go into Tamash. We'll continue to really bring that power point down. From a growth standpoint, it's clear that ultra thins and tablets will be two of the biggest growth vectors in the mobile space. We're going to go after that with a vengeance.
What we see there is taking all of the low power techniques that we know about, we can absolutely get x86 to less than 2 watts. If you think about that capability, you can take the full Windows 8 capability into an x86 processor in a tablet system that lasts for over 30 days battery life. That is very much within the realm of our capability. Now when we leverage the APUs with more graphics and video processing and all that stuff, you get the differentiation for our product set. With flexible SoC designs, what you're going to see from us is more products faster in an ambidextrous way so that we leverage the best of all ecosystems. As you think about Windows or Android and other areas, clearly all of those ecosystems are important for growth in this market. We will support those.
We are really focused on how do we differentiate in the market. There are a lot of players who are out there talking about tablets. Frankly, the volume isn't so high, except for the first one. When you think about how you differentiate, though, it is differentiation on the user experience. That is where I think we are unique. It's not just about the processing technology. It's about the user experience. It's about how we can manage and really marry that software capability with the hardware capability. You'll see a lot more focus on how do we bring those special features to the customer. Things like 3D content, things like biometrics and facial recognition and speech recognition, all of those are uniquely accelerated by the APUs. We can do that at very low power with our capabilities. The server market is also a place where we have tremendous opportunity.
Here, I'd like you to think about it in a slightly different way. We've talked about our baseline server roadmap. We are absolutely committed to that roadmap going forward with Interlagos and Abu Dhabi. What we see is an opportunity to really change the game. When you look at the server market, those top levels, that cloud computing, that mega data center market is really changing server dynamics. Now you're not just talking about racks and racks of servers that just care about power and performance. You're talking about specialty workloads. It actually will fragment the server market a bit. You will see different solutions that will fit different server workloads. It's not a one size fits all. That plays into our strength. Once again, we can optimize for different workloads with our acceleration capability. There are some workloads that will be accelerated by the APU.
We're going to use the APU in those workloads. There are some other workloads that may not be so accelerated by the APU, but are accelerated by more lots of little cores. Lots of little cores in parallel is another option that we have. Our goal in the server space is to really attack those disruptive workloads with all of our capabilities, including the SoC infrastructure that Mark talked about and the ambidextrous solutions to look across various ecosystems. This is where it's about system level optimization, not single threaded performance. We'll really be able to take advantage of that in terms of growing our server market share and also from the standpoint of really growing our presence in this new and emerging market. I'll spend a couple of minutes on embedded. The embedded market is one that I know very well, given some of my history.
The embedded market is really a market of lots of different subsegments. When I think about embedded for AMD, I think about embedded as let's take all of the technology that we have, the basic processor technology, the basic APU technology, the basic graphics acceleration technology, and really pick the subsegments that are going to benefit from that. That is really the key, taking our technology into different markets. If you think about where that is, there are a lot of APU friendly segments. I call them APU friendly because they can really benefit from the graphics and the parallel processing acceleration capability. Things like digital signage, things like thin client, things like medical imaging, when you think about how those will look, gaming is a great place for our APU technology. There are some subsegments of communications and storage. This is not necessarily about new product SKUs.
This is about taking our current product SKUs and getting really clear in terms of how we go to market with them. How do we grow the embedded ecosystem around AMD? This is an area that we haven't spent time on. This is an area that we haven't invested in. This is an area that we believe we can really add value because of the capabilities that we bring in terms of processing. With that, I feel like I'm now a veteran of AMD at 30 days in my first financial analyst day. I hope what I've given you today is, one, we wanted to make sure that we shared with you all the product details. There were a lot of questions about what we're doing with our product roadmap. I hope what you take from it is this product roadmap is really good. It's really good.
It's really focused on execution. It's very clear what we have to do over the next couple of years. That clarity is extremely important to ourselves, to the marketplace, to our customers. That is very clear. What I also hope you get is a little bit of a feeling of how much excitement we have and how much excitement that I personally have. I look at AMD as an opportunity with tremendous technology potential. It's like every engineer's dream is to have all of these bits and pieces to play with, but really putting it together into products that will grow market share, that will return sustainable value, and that will really touch the consumers from a user experience standpoint. That's what we're really looking forward to, really skating to the puck for the future. Thank you very much.
With that, I get to introduce Thomas Seifert, our veteran of our crew, who can take us through the financials.
Thank you, Lisa. It's a delight to be here today and talk to you. Who would have thought that two or three financial analyst days make me a veteran? Quite a surprise. I also have to tell you that from a CFO perspective, the last couple of months have been a sheer delight for me. I have colleagues, new colleagues now that not only do they know numbers and understand numbers, they actually enjoy working with numbers. You feel the excitement, hopefully, of the presentations today. I feel the excitement because the numbers that I look at are shared, and the value that they represent are appreciated by my colleagues. We have a good time. I'm excited about what is happening. Let's use this now. Let's put some numbers around the things that you have heard. I'm going to take you through a couple of things today.
I would like to have a short recap of 2011, talk about the business model we have put in place, what went well, where we have done well, and where there is room for improvement. I'd like to talk to you about our capital structure. We've made significant improvements over the last years. We've done tremendous steps forward in 2011. Where's the room to do more goodness? Where is the optimal capital structure for an AMD? How do we look at that? Of course, I would like to take you through our 2012 operational model. You've heard a lot of things that are going to change, exciting things, how we adjust our business model, how we reposition the company, and really make aggressive use of this fabulous model that we've ended up with.
Do this in the confines of the financial parameters that we have set for ourselves and make sure that we keep the financial momentum that we have started to build. Also give you a glimpse into what is beyond 2012. What is the long-term financial model that we can imagine? Where can we take this business model from an earnings perspective? Let's talk about 2011 first. We've executed well to our business model. It was a year where we launched APUs. We shipped 30 million. We shipped also 100 million DX11 devices on the graphics side. We launched the Brazos APUs, the low power devices, a very successful launch. We've left some opportunities on the table, too, especially in the server space. We picked up speed in the second and the third and fourth quarter, but we left some opportunity on the table.
This is room for us for this year. I think we clearly stabilized the financial performance. We found a way to improve how we run the company around working capital. It had huge benefits, as you will see in a second, on our cash flow generating capability. I think when we started the transition into the fabless model, we said the real strength of AMD in this new setup is going to be to transform ourselves into an IP generating and free cash flow generating engine. You've heard from Mark and Lisa and Rory this morning what we've embarked on in terms of making this IP generation come true. I think the progress you can see over the last two years turning AMD into a free cash flow generating engine has made tremendous progress. I'm going to talk about this. Of course, we've shown consistent profitability for the year.
From a revenue perspective, last year was certainly a notebook year. We launched exciting products. We dominated U.S. retail in the fourth quarter. I think five out of the six best selling SKUs were AMD-based notebook SKUs. We dominated Black Friday, 70%- 80% of the revenue was on AMD-based systems. We've made tremendous progress in the emerging markets. We've grown market share in China retail for three consecutive quarters in double digit numbers, tremendous progress. We left opportunity on the table on the server side without any doubt. We picked up speed with the rollout of the Bulldozer-based products in the second, third, and fourth quarter. There's opportunity, tremendous opportunity moving into this year and going beyond exploiting that. Good progress. How does this help us with our financial performance?
Just to remind you what the targets were for 2011, we gave what we thought at that point were ambitious targets in terms of gross margin, in terms of operating expenses, in terms of operating income, and also in terms of free cash flow. I think we executed well to the targets we set. We came in the range of the gross margin guidance we had, despite some of the challenges and despite the fact that we left opportunity on the table in the server segment. I think as a company, we did extremely well, staying very disciplined in our approach how we manage operating expenses. For that very reason, we landed well in the OpEx corridors in terms of R&D and SG&A that we set for ourselves. We generated operating income. Most of all, we made tremendous progress on the free cash flow generating side.
$528 million of free cash flow last year was a tremendous step forward from $355 million to $528 million. If you look at our free cash flow yield, I think this puts us at the top of our peer group. This is an important fact because this allows us for this year and moving forward tremendous amounts of operational flexibility to further work on our balance sheet, but also to make sure that we have the strategic flexibility in order to improve the business model. With this in mind, what did we do with all this good free cash flow? How did we progress on our balance sheet side? What did we do well, and where's the room to move forward? Mr.
Chart, with all the changes you have seen and all the changes you have been indicating in terms of how we transform ourselves into an IP generating engine, it's really important to understand that as a company, we are committed to real fundamental changes. You heard us talk a lot last year that we started to redesign the core processes in the company, how we go to market, how we run our supply chain, and how we design our products. Why is this important? It's important for really three reasons. If we want to make this shift in our business model, become a much more aggressive player with our fabless business model, we have to increase the core speed of the company. In order to achieve that, you have to redesign the processes that determine speed within the company.
That is how we translate market needs into products, how we design products. You heard Mark talk a lot about how we make changes there, and how we run a supply chain. An SoC model is going to be a challenging model in terms of how you place inventory, how you manufacture. We have to be prepared for making these changes. We think there's about a $700 million productivity potential that we have attacked in terms of lower working capital, in terms of additional revenue, in terms of being faster, in terms of time-to-market. It's goodness, and it's important work that prepares us for making this shift on the IP side. You also have to look at the restructuring that we have done last year from that perspective. This was about making the organization faster and more agile.
We used this opportunity to increase the span of controls, to take layers out in the organization, to bring this agility into the organization. It's not so much about saving this money. It's about reinvesting this money. About $160 million out of the savings we have will go towards investments in the areas you heard Rory, Mark, and Lisa talking about, into consumerization, into cloud applications, and in device convergence, making sure that we are well prepared for this. What did we do with all this good cash flow? What progress did we make on our balance sheet over the last 12 months? I think it's clear that we increased liquidity. We increased our liquidity now to $1.9 billion by $125 million. We worked hard on further debt reduction. We took out $200 million from the debt we had. We brought down our leverage in a very significant way.
We maintained our credit ratings. You actually saw that two days ago, we were put on positive credit watch from S&P. It's a very important step when I talk to you about our optimal capital structure. Good progress on the balance sheet side. With this debt reduction process, we came from close to $4 billion not too long ago now to about $2 billion. Most important, not only did we reduce the debt, we brought down the maturities in very sizable chunks. It's our target to keep the towers in the range of $500 million, so they're always in striking distance of the annual free cash flow capability that the company has. It's a very important step for us, and there is room to go. What room do we have? We did well in 2011.
There are $487 million of our five and three quarter maturities going to come due in the summer. If we were to take them out completely, it would help us tremendously on our leverage. It would help us tremendously taking down interest expenses by about $28 million, and it would have earnings per share leverage of about $0.04. There's significant leverage opportunity left on this balance sheet. Coupled with the free cash flow generation capability the company has now put in place, I think that is really good news from a balance sheet and also from an earnings perspective. If this is all true, where do we think the long-term optimal capital structure is for the company? We all know the semiconductor environment is a volatile place. There's a lot of uncertainty in the capital markets today.
If you keep this in mind, put it all together, we think that maintaining a cash level of $1.5 billion is going to be a good neighborhood for us, a good range for us. We also want to make sure that we keep the debt towers below $500 million. They are in the striking distance of what we can generate as a free cash flow capability on an annual basis. We will continue to deploy cash to reduce that opportunistically. We have line of sight to becoming net cash positive. That is the near-term goal to achieve that as fast as we can and get the leverage, the debt to EBITDA leverage down to a factor of 1.0. We also work hard on our investment, on our credit rating.
With the last move we saw from S&P, I think a good target for us is to reach really investment grade level. With the work, this is not going to happen overnight. If we continue the progress we have been making, if we continue to work hard bringing the debt down, I think this is a good target to have for us. Let's talk about 2012. You heard Rory talk about and Mark and Lisa about the changes that we are going to implement, really taking an aggressive advantage of this fabless business model we have, taking this rich IP that we have and put it into marketplace in a better way and new products.
We have to make sure, this is my job, that this all happens in the confines of the goals we have been setting for us, that we keep the financial parameters and the financial momentum we have been generating over the last two years and maintain this momentum. I think with all the advanced work we have done in terms of really restructuring the company, redesigning the processes, I hope you'll see that we are able to achieve that in a very good way. Before we go into the details of the 2012 guidance, I would like to give you some key drivers and color around key drivers that are good to keep in mind. If you look at this year for us from a revenue perspective, without any doubt, there's a lot of tailwind we are going to get bringing differentiated APU offerings in the market.
You heard Lisa's excitement when she talked about Trinity. The design win momentum is higher than we had on Llano a year ago. We feel good about the momentum we have been generating with Brazos in the low power space. This is going to continue with Brazos 2.0. I think it was clear that there's room left in the server space to improve. This would help us also in terms of product mix and ASP performance. There's some headwind, to be very honest, from a supply chain perspective. We all know the uncertainty around the hard disk drive space. This will be something that we have to deal with, especially in the first and second quarter. While we have been making extremely good progress on our wafer supply agreement, transitioning from one wafer supply agreement to the next will also put some constraints around it.
We are going to continue extremely hard on the earnings side to manage operating expenses well and stay disciplined. I hope I can impress this upon you on the next chart. From a cash flow side, if those things come together, we manage our inventory well. We have redesigned our supply chain to make sure that we can take working capital out of the system. We watch this very carefully. We believe that we will continue to see very strong free cash flow momentum in 2012 moving forward. How does this all add up now? What is the take on 2012? We expect the market to grow about 5% this year. Lots of numbers whether what the impacts are going to do on the supply chain side, 5% across the markets. We plan to outgrow the overall market growth.
We think we can take gross margins in a range of 44%- 48%. We started off well this year. We gave guidance for 45% for the first quarter. We are going to build on this. We have started well, and we build on this momentum. We see, as I said, tailwinds, especially when we go in the second half of the year in terms of product mix, productivity, on ramping the new products. Gaming revenue, you will see this in a second, is going to help us. We see we're going to see some headwinds. There's uncertainty, as I said, around the supply chain, the hard disk situation. We will have yield challenges because there are a lot of 28 mm products, especially on the graphics side, ramping into volume. We're going to manage operating expenses very tightly.
This will be the first year where we really give hard operating expense guidance, no ranges, $610 million per quarter on a yearly basis. We started off low this year. We gave guidance for $590 million for the first quarter. This gives you an indication on how this should be distributed over the year. We will see some climbing up towards the end of the year as we get a lot of good products ready for 2013 from an engineering and tapeout capability. Overall, we are going to manage operating expenses as tight as we have managed them last year. This will give us positive operating income. Taxes, you should model around $3 million per quarter. Capital expenditures in the range of $200 million.
That is really very much backend related, making sure our packaging capabilities are where they need to be, especially with so much of our product moving into the low power and ultra low power space. We have certain investments also in our own data centers moving forward. Overall, I think it's good guidance for 2012 from a revenue perspective. As I said, we will continue to grow and gain market share. You heard Lisa, you heard her strong commitment in the past she got from Rory, take this IP and turn it into market share gain. Strong growth on the notebook side, serve lots of opportunity. We feel good for this year with the momentum and the trajectory we took out of last year in the third and fourth quarter. We will see good gaming revenue opportunity, especially towards the end of the year.
From an overall growth perspective, 2012 should be a good year for us. All the changes, all the indications you have seen, how we move ourselves from an agile to an agile IP generating SoC-based company, taking really all the richness we have in our portfolio to the market. Where does this constrain us? Where do we see the long-range operating point? We think we have line of sight to a target of gross margin beyond 50% moving out of 2012. The key drivers are growth acceleration. You heard Lisa talk about the market growth opportunities we have in 2013 and 2014 and 2015. The enhanced product mix this is going to drive, also in adjacent market segments that we have not been serving today, embedded for one, server for others, also taking us down into the ultra low power space.
If one message should have come across today, it's about disciplined execution. We have seen tremendous tick up of our capability over the last five to six months since Rory joined. There is a lot to go, a long way to go. We're making huge progress. This will have an impact on the earnings power of the company. These will be the key drivers. You saw the key changes that Lisa talked about, how we reposition ourselves in the cloud space, in the consumers, what impact the consumerization is going to have on our product portfolio, and how big conversion is going to play for us. These are the key changes, line of sight to a gross margin beyond 50%. You marry that with the free cash flow capability that we have already started to see. I think this is a good and exciting outlook moving forward.
With this, let me really summarize it. We have seen that the financials are stable. There is significant opportunity to grow for us. Lisa has been very clear about this. Rory has been very clear about this. It means that with this growth, with the performance we have implemented, our cash flow generating ability will continue. This will give us the flexibility we need from an operational perspective to continue to work on our balance sheet, to make sure that we really lift the potential that has in terms of earnings per share impact, but also provide us the strategic flexibility that we need in order to move forward and expand building this business model. It allows us, of course, also to significantly reduce debt.
I think it allows us to make this year a good year in terms of repositioning us, not only as a cash flow generating engine, but making and transforming us into an aggressive, fabless IP generating play. With this, my part is done. I think we transition to the question and answer session. Thank you. Thank you.
OK. Now we're going to kick off with the question and answer session. We have some roaming mics around the room. We'd appreciate you raising your hands, which you're already doing. We will start the QA session. I just ask before we start that you ensure you speak into the microphone with your question so the folks in the webcast can engage. We have our first question over here, please.
Hi. Thank you. Lots of energy here today. I was curious in terms of some of the new programs that you're starting, and Lisa, your comments on doing more faster. It seems like your OpEx budget this next year is going to be roughly flat with what you had in calendar 2011. How come there's no impact on the R&D side to doing more faster going forward? I was wondering if you could share this because we didn't or I didn't catch the desktop commentary in terms of the roadmap or what your view on desktop or are you lowering R&D in that part of your roadmap? Is that why it's coming down or just some clarity there? Thank you.
Thomas, would you like to?
Yeah, let me take the operating expense question first. That's a very valid point. That's why it was so important for us to start this project of re-engineering the company and the core processes first. That allowed us to lift a significant productivity potential in the company by being faster to market, by taking working capital down, readjusting our R&D processes moving forward, how we develop products. You have to keep in mind that the restructuring we have done, delayering the organization, making sure that we have the right span of controls, not only leads to a faster decision-making organization, a faster reacting organization, but it also provided us with a $200 million spend envelope that is not going again into rebuilding SG&A. It's going to be reinvested in R&D.
From year- over- year, if you include this reinvestment and the productivity gains we have made, I think we spent more money. We put more resources towards product development on the R&D side. We've been able to contain it within a very tight budget. I think that was a good achievement for the AMD organization worldwide.
Yeah, no doubt the productivity gains that we got for Project Win and the process work are now flowing through and moving through the system. The work that Mark is doing on the architecture work, the reusable, again getting more efficient, the common best practice tools, and then the work that we're doing in SG&A around the right kinds of marketing investments. We must make sure the execution engine can deliver. If we spend additional marketing dollars too early, it won't produce the outcome that we want. We've also done some work around the coverage model to make sure we've redistributed our coverage structure to the high growth markets that allows us to capture some of that emerging growth, all with the focus to allow us to invest into those attack areas that we see as the opportunity moving forward.
Maybe I'll just comment on the desktop comments. Certainly, we believe that APUs are going to be our growth vector, and that's where we're focusing the majority of our investment. The desktop high end will leverage off of the low end server, and we'll have a Piledriver version as well on top of the current roadmap.
Grace, I believe there was a question over here. David, just over here, please. Thanks.
It is on this again.
Thank you. David Wang, Wells Fargo. Can you tell us what the minimum threshold size for servers needs to be for you to be able to justify investment in servers? What market share do you need to get to from where you are currently on server processors?
Yeah, I mean, clearly, we view server as a large opportunity for us from where we are today. Our aspiration is to be double digit market share. Of course, as soon as possible, like my boss would say. The key is consistent investment with our customers to be double digit market share. Of course, as soon as possible, like my boss would say. The key is consistent investment with our customers and consistent investment in the infrastructure. That is where we're going with servers.
We've seen two solid quarters of improving velocity in the server business. We've got to build on that quarter in and quarter out, each quarter adding additional market share. It's about creating the workload focus, the relationship, and then the consistency of delivery, particularly in the roadmap, as well as on the supply chain that allows us to build that trust again for the future growth. Definitely, double digit is where we want to go over time.
Great, hear this gentleman here.
Thank you for taking my question. You guys really didn't talk much about manufacturing. I know that's been one of the challenges in 2011, is your dependence upon GlobalFoundries and their ability or lack of ability to produce products. I guess what's the strategy going forward there? Where are you with other foundry partners on the processor side? Can you help give us some confidence that you'll be able to churn out some of these Trinity APUs? Thanks.
Yeah, good question. You probably saw that we have one breakout session prepared afterwards that is talking about our supply chain approach moving forward. You will get more granularity there. I would say we have made tremendous progress in the third and especially in the fourth quarter. We have been rather explicit on this also in our earnings call. We put as partners together with IBM and GlobalFoundries and us significant resources into Tracedon. We have been doubling out. We have been increasing output on Llano 32 nanometer-based products by about 80% quarter- over- quarter. The progress is good. The status on Trinity is really good. You heard Lisa say that we already ship products to customers. That is a very encouraging topic. Partnerships are defined that you work together not only in good times, but also in difficult times.
I think we have brought the organizations from the two or three companies together in a very good way. We feel good about the progress we have been making. I think this goodness in terms of development on the relationship side is also what is driving now the way for supply discussions. We said on the earnings call we're on the finishing road, finishing last mile. We're confident that we bring this to conclusion in the not too distant future.
Let me add a comment as well from how we're approaching this from the development side. To have a sustaining deep relationship with the foundries is essential. What we've done is put a very rigorous process in place. It starts early. It starts marrying their design elements with the foundry capabilities, having very clear communication and partnership with the foundries, and mapping that forward, checking it along the way on both the design and the foundry side to ensure that we're marching in unison. That's what we put in place going forward, and it's proving very effective.
There is a deep focus on building a true partnership. This has been an intense amount of time that we've focused on over the past five plus months to really create that partnership, rebuild it. I can tell you firsthand that the relationships we have with the foundry partners, Global Foundries, ATEC, other ones across the planet, we are developing a much better relationship. As Mark pointed out, it's about the teamwork to solve the problems. We win together. We have a partnership, like Thomas said, in difficult times and in good times. What we are seeing is the focus on execution, running the test chips through the line, the gathering of data, the movement with John Docherty's team in supply chain, working with partners from IBM and Global Foundries. We're seeing real focus on day in and day out executional improvements.
Because of the work that we're doing at the partnership level, we're getting the right kind of uptick from their side of the organization as well. There is a commitment to this partnership. You can see we shipped 80% more 32 nm product in the fourth quarter. We made our customer commitments. I'd like to improve the SKU. It was still a little too far back and loaded. I want to make sure we deliver every customer commitment within that month this quarter. The quarter after that, within the week that the commitment is made, week in and week out, those customers count on it and know that it's there. You should know that partnership is improving.
Great. There's a question right back to the room here.
Sure. Let me just start off with Thomas and probably Rory on pricing. What is your pricing philosophy going forward? Do you think there's enough differentiation that you offer for you to move up the price stack? Conversely, to ask it differently, do you think there's enough flexibility in your model today to absorb any type of price pressure that could come in the industry as you fight not just Intel, but also the ARM camp?
I'll take the first part on that. From a business unit perspective, and perhaps Lisa wants to comment on it, we have to focus on the four P's around the way we run that, primary focus on the customer with the business units, driving the market analysis, where the market is going. What that means is use all four P's. Do the analysis on the product. How competitive it is? Where does it sit? When you're talking about the fastest processing capability on the planet with a 970 graphics card, you are going to have a pricing strategy that's different. You might run it at a PFV of, I don't know, 1.05 or 1.08.
What we want to do is create the analysis across all four P's: product, placement, promotion, price, and make sure that we have the competitive analysis and then the discipline at the BU structure to run that. As sales executes within the quarter, they work within the framework of that. We also want
To move the skewing of the product set so that it's not as back-end loaded, I want to make sure that we're selling product and delivering it on a more consistent, linear basis. We want to make sure that our promotions line up to, in the emerging market, things like Chinese New Year, you know, like Black Friday. We create that kind of discipline and structure into our execution models around true pricing analysis and competitive analysis. We'll get better results.
Maybe I'll just add to that. I mean, I think it's clear that, you know, our success is about a differentiated value proposition. It can't just be about price. It's about bringing in the right feature set. It's bringing in low power at the right places. It's really tailoring to the solutions that are needed in the marketplace. I think that's what we have with all of our IP, really putting that together. You're going to see a lot more focus on market-centric value proposition, ensuring that we have the right SKUs for the right geographies as well to really broaden that out.
Yeah, no, go ahead, Uche.
Sure. Can I just ask a follow-up on the Global Foundries answer you were providing earlier? Is it possible for us to really understand in a little bit more detail what happened between September and December? As of September, the conference call was not so good. By December, things had improved dramatically. Is that something that is sustainable? Also, to what extent will that affect the wafer supply agreements you have with Global Foundries? Are you going to move back to buying wafers now, or are you still going to keep buying good guys out? Thank you.
Want to start?
You can go first.
I will go first. We put in a lot of work together, and we certainly think that the efforts that resulted in this significant improvement are sustainable. The way how we have started into the first quarter makes us believe that we are in a really good trend. Semiconductor manufacturing is not a trivial thing, you know this. I think the improvement has been in a very systematic way, how we communicate, how we look at the problem, how we look at data, how we run the manufacturing line, how we communicate between the entities. I think this is a fundamental improvement. For that very reason, it is a sustainable improvement. That is how we look at it. Of course, you know, this significantly improves emotions in a negotiation situation. We will talk about the specifics on the wafer supply agreement when we have signed it.
It is premature to do this today. As we said before, we have made significant progress. I think we have lined on site to close it in the not-too-distant future. At that point, we will talk about the content, not before.
It's fundamental in terms of, as you focus on that execution from back in August and September through December, it's rebuilding the relationship across all levels of the organization, getting into this idea about ownership and commitment. People think of it as a soft kind of concept, but it's this idea that a commitment and ownership are linked together. It means that when we make a commitment to our partners to deliver a certain set of supply, we have to deliver it. It doesn't matter what part of the model or the process breaks down. We own it end to end. It's sort of like the movie Apollo 13. Those guys had that capsule, had the problem, and it blew up there. They had to figure out how to get it home.
It wasn't about this team let us down and I was good and the rest of them were bad. It was like, no, we have a commitment to our customers. That commitment I own. We have to own it across organizations, across partners, and we have to deliver it. We've systematically implemented the changes to put in place the test chips through the line, gather the data, systematically do the Pareto breakdowns of each of the defects, do the analysis on what we think the return. Now we're seeing a very consistent level of execution in terms of our forecasted curves.
Yeah, I can't emphasize enough what Rory just said. I've got many years of deep fab relationships. It's very clear the equation you need. It's exactly as Rory described. It's very candid. It's based on goal alignment and rigorous data and execution. That's what we have in place at this time.
Great. There's a question just here, this gentleman. He's had his hand up for a while. Thank you. I'm then behind him afterwards.
I have a two-part question from a list of many, so I'll keep it short. The SOI, I'd like to know what you're going to do with the SOI because it's so very connected to low power. The other part is your manufacturing strategy, whether between Global Foundries and whether or not there's going to be a TSMC mix considerations.
Sure, you take it.
Let me take the first part in terms of SOI. If you look at our product mix today, we have 40 nm on bulk. We have 32 nanometer on SOI. We're going in 28 nm bulk. We're determining beyond that the next technology node because what we're focused on in each of these is it's about optimizing our design for low power and performance per watt, as we described earlier. It's the whole combination of the elements of what we're doing from a design, how we're managing power, what we're doing across the platform, and of course, that core technology is very important. It's about putting all that together, generation by generation, node after node, and picking that best combination.
Our roadmap through 28 is clear, and we're looking beyond that as to what is that best technology that marries with the rest of our design and elements across the stack, which we're doing to have leadership performance per watt.
The gentleman just in front of the guy just asked a question.
No, we didn't get the second one.
You guys want to start?
Yeah, I said, no, we have two foundry providers today. Our graphic part, it's coming from TSMC, also from Lisa that we've been making good progress. We launched our 28 nm product successfully last year, and we're going to extend that product family across the whole product stack over the course of this quarter, actually. Very good progress. We will have a balanced strategy moving forward. We're not going to give details on which product at this point in time is going to be manufactured where.
Okay.
Thank you. You talked a lot about SOCs predicted during the first half of this session today. I guess I'm a little confused in terms of where you actually see the company as a whole going along this route as we move out. I mean, are you seeing yourselves eventually becoming an IP factory that's selling SoCs, or are you becoming an SoC integration factory that's just handling the integration and the interface with the foundry? Where do you see this going? How is that actually creating value for you and your customers? I suppose finally, why do you think you're going to be better at doing SoCs than the Qualcomms, the TIs, the NVIDIAs, or frankly even the Intels over the timeframe that you're talking about?
That's a great question. Let me take the first stab at this. It's really fundamentally, this capability that we described to you here today provides a multitude of benefits. The first and foremost of why we're going to this approach is that it speeds our delivery. It's internally focused. We have a set of platforms that we've been targeting that we want to speed our delivery of our solutions into the marketplace. First and foremost, it's about execution efficiency because it drives as you architect that IP. I called it the treasure trove of IP that we have here at AMD. As you architect that to be a clearly defined IP block with clear and rigorous interfaces of how those IP blocks work with the rest of our AMD IP, it speeds our time to market. That's first and foremost the benefit.
It just adds up from there because, again, it opens up the flexibility for us to even more tailor solutions to our customer needs, gives us that flexibility to bring in third-party IP, third-party ecosystems that are out there so we can have better solutions to the market. To be honest with you, once you've done all that and you've sped that ability to meet market needs, you've frankly taken your IP and you've architected such that it creates yet further monetization opportunity. If you chose, in certain cases, to license that IP, we would have that flexibility as well. It's an additive benefit. First and foremost, it's about allowing us to be flexible and quick at getting our IP to market to meet those customer needs. I look out at the competition. There's a whole range and degree of how this is implemented.
What we're doing is what we believe is a best-of-breed implementation. It's about rigorously applying a methodology and a set of interfaces that speeds that IP to market. That's what we're doing.
I want to build on that a little bit before we go deeper into it. Think about what Lisa talked about in our core businesses, right? She talked about graphics, she talked about clients, she talked about server. That's our core space. We build solution platform solutions that integrate software, they integrate graphics, CPUs into a solution that allows players to play in that space and win, whether it's the OEM or the ultimate customer. What we want to do is we're going to focus and stick to our knitting in that segment. That's the core business. That's in the tactical timeframe, 2012, 2013, et cetera.
As we architect that architecture and we become more agile and we create a standard architecture set from, say, 50 millimeter chips, 100, 125, 150, 250, whatever, that architecture then we be able to deliver it more efficiently, create more value on it, integrate more to get better low power and more efficiency. If there's adjacent spaces, game consoles, into perhaps medical or communications that are very much aligned around the model that we're using, we can expand the TAM and also leverage that IP on a broader set of solutions. Make no mistake that the focus and the core is to use this architecture to win in the markets that we're there. Lisa, did you want to add anything?
Yeah, maybe I'll just add because I think your question was why do we think we can differentiate in the space? The differentiation is the IP. I mean, the IP is tremendous. The SoC methodology is a way for us to get faster products to market. It's really a way to reuse all that IP in a more efficient manner. It's not like a trade-off of, hey, you have to have an SoC factory or an IP factory. It's really trying to marry those two things together to get more products to market faster.
Do you have the SoC expertise in-house yet, or is this what you're hiring? Because it looked like on your product roadmap you won't really have the full capability to do this till 2014 and beyond, which is an eternity in this space. Is there a danger that if it takes you that long to get to the point where your time-to-market is quicker than it is now, quicker than the competition, that the market sort of moves out from under you in the meantime?
Yeah, no, we're actually doing this as we speak today. If you think about our 2013 roadmap, the Kabini and Temash points, those are full system on chips. A single chip solution, we've integrated the IO on there, and we'll do more as we go into the future. This is not a, you know, we're going to wait until 2014. The SoC expertise is in-house. It's re-architecting the solution such that we get it into the products in a timely manner.
That is part of the reason what drew us all toward AMD was that capability, the kind of engineering capability, the kinds of people that we have here that have this almost entrepreneurial spirit. They've got all the background. As these inflection points occur, we're really trying to marry that capability where the market's going. It's not that we're going out and hunting. Sure, we'll look for some pieces of IP to add to the portfolio that builds out our capability to answer that solution. That is what brought us all to AMD because the engineering capability, the talent, the piece parts are here. Now it's to capture this wave and weave them together in that model to win.
Great, thank you. Next question. I'm here in the middle of all this junk on here. Thank you.
Hey, thanks a lot. Thomas, in your 2012 outlook, you highlighted ASPs as a driver for both gross margin and earnings. ASPs, from what I understand, have been fairly stable the last couple of years. Could you be more specific on how you see it trending in 2012?
Yeah, for us, ASP performance this year is very much a matter of product mix improvement. I think it's without any doubt clear that we have room to grow in the service segment. This will have a potentially significant tailwind for the overall ASP performance of the company. We think that the launch of the Trinity platforms, bringing this into the ultra-thin segment, is going to help us tremendously improve the product mix and, with that, price ASP performance. We also have opportunities towards the end of the year when we think we will also see strengths in the gaming segment.
Would you say that ASPs have an upward bias in 2012?
I would say there is opportunity for that, yes, especially in the second half of the year.
Great. We have a question here on this side in the middle. Thank you.
Thank you. A couple of questions. Lisa, you said your goal is to get the ultra-thins to about $600 price points. I'm just curious, you know, what needs to happen for the pricing to come down so much? Is it as simple as just, you know, offering a cheaper CPU, or is there anything else that you can do to lower the overall pricing, bill of materials pricing? The second question maybe for Thomas. Thomas, you said in your forecast you expect the GPU market to grow at about 5% this year. Given the optimism about the APUs and ultra-thins, what gives you confidence that, you know, that market will even grow this year? Thank you.
Yeah, the first part of that question, I think we see tremendous value in what Trinity is able to give us. If you look at the power performance characteristics, the fact that we can double the performance of Llano in performance per watt of Llano and get into a 17 W price point. We've been working with some of the Taiwanese ODMs to really put together reference designs. The one that I showed you was from Compal, but we've been working with a number of them to really get the optimized footprint. The Z footprint is by going to a BGA solution and then really getting the overall bill of materials cost optimized. I think there's a lot of anticipation and belief that once we can get the ultra-thins down below, you know, these higher $1,000 price points, we can really drive volume in that area.
That's the reason that, you know, Trinity has so much capability and we're putting so much into the work with the ODMs and OEMs.
You know, I also spent a lot of time at an OEM over the past five years. If you look at what drives the whole cost of the end solution, there's really only five parts, and it's disproportionately slanted to two, operating system and processor in terms of that cost. We believe with the processing capability, the experience you can create with the Trinity APU based on the workloads, just like we did on Brazos, we'll be able to introduce that solution at a more effective, valuable price point. What's important that we want to share with the OEMs is you can use that processing capability and you can redistribute the profit pool in the model where it is today.
We think with Trinity, that can actually help them differentiate, but also redistribute in terms of those five major product sets within the device, part sets where the profit is generated. We think that doesn't always have to be at the lowest price cell. We can use Trinity across a series of price cells and help move around that profit pool. That's very attractive to OEMs.
With respect to market growth, I think it's a good question. There's considerable debate on where the unit volume is going to go. We used, you know, an external reference, market research reference that we think makes sense at this time. I think there you can put a lot of arguments around it either way. Where are attach rates going to take us? How is the market going to develop? What impact are going to APUs going to have on discrete level of traffic? From the visibility we have today, I think that is a number that makes sense, and that's why we have used it.
Just here because there's a lady here. Thank you.
This is Joanne Feeney from Longbow Research. You spoke earlier about the old AMD way of trying to do extra design complexity to make up for a process lag. How does your decision to, in some sense, abandon that, or at least lessen the importance of that, mesh with the need you have to continue to push for the lowest possible power and highest performance in the server space? I would imagine there would be spillovers from those innovations back down to the SoC space, but perhaps you could enlighten us on what kind of lag that would be and how much spillover you do anticipate.
Sure. I can take the first part of that, and that is when you look at what we're doing across low power, the technology semiconductor node is certainly a piece of this. What you're seeing as we piece these solutions together across the different elements of IP, we're driving performance higher by bringing this together. When you look at the very basic premise of the APU, bringing the CPU and GPU together, and you saw on the slides, for instance, that Lisa showed, the change in slope as we bring the different IP elements together, it's significant. When I talked about the elements we're putting into the design for power management, it turns out that those elements have dwarfed the power savings that we get from technology node. Technology node remains very important.
If you look at, I'll say, the piece of the pie in terms of giving us the overall performance per watt, that overall experience that we can deliver to the customer, it turns out it's a smaller piece of the pie moving forward than it has been. Remains very important. We focus on it. We talked earlier about the partnership that we have with our fabs to make sure that we're optimized around each fab, we're optimized around each transition. It's about bringing the total solution together. That's what we've done successfully with Brazos. That's what we're going to continue to do. Lisa, do you have any comments?
Yeah, specifically around the server market, I think it's very clear that you really have to choose where you're optimizing your portfolio. If we're going after frequency and single-threaded performance, yes, you got to be on bleeding-edge process technology. If you're going after these new workloads, which are really much more tailored and can benefit from acceleration and benefit from everything we can bring with heterogeneous computing, it's a lot more about architecture than it is about process technology. I think we want to marry the drive for performance.
We're continuing to invest in the server cores, and you saw our investments out to the excavator generation, but really marrying that system-level view of how do we get the total cost of ownership, how do we get the new workloads, how do we use multi-threading, how do we work with the software and the tool vendors to make sure that we're taking advantage of all of that to give us a server performance advantage.
If you look at the work that Chekib and his team are doing in the disruptive space, now we're looking again even further into that workload variation that's already occurring and going to accelerate. The dense type of solution becomes a very different power envelope and target kind of solution in the way we stitch it together with the fabric. We think that's going to be disruptive to the old models that have been in place for some period of time and why we're playing in that. We're at a play, in terms of leveraging the CAT cores and the dense kinds of servers, as well as on the traditional evolutions of Interlagos into Abu Dhabi. We think those are the right game plans and to really target it at a workload level. That's really kind of the focus. It's about the experience at the glass for the client.
How do we integrate it? What's the graphics performance? At the server, it's around the workload. We don't want a one size fits all. For us to win and for us to create value for our customers is to really focus on creating that differentiated experience like we did on Brazos, like we're doing with Interlagos and the work that we're doing in some of the traditional workloads. Where we're going in the disruptive place, again, tailored to the unique experiences and workload, we think there's a huge amount of scale and they're at the beginnings of that tipping point that move in those directions. That's how we're trying to skate to the puck.
Great. Do you have a follow-up, JoAnne?
Yeah, just the theme of tailoring has been very prominent today. There are some concerns that that means you're going to be going for a higher mix, lower volume, you know, set of output. How do you plan to manage that risk?
Sure. As I said in the beginning, there's two lifts to create value at AMD. The first lift is around execution, creating an execution engine that can deliver and really drive on its base. This business that we have is a good business. If we execute better on it and drive scale, we have a high-scale client mobility business. We can continue to extend that. We can build the level of the foundation of trust across partners, suppliers, et cetera, that allows us to grow that base. As we move forward, what we really want to do is make sure, where was I going on that one? I'm going to do a Texas thing. There were three things, but I'm only going to talk about two of them.
It makes a lot of sense.
What was the question?
The purpose-built.
Oh, and the purpose build. How not to get distracted. Thanks. He didn't get that help in Texas. I'm from Austin, so I'm sorry about that. Think about it from the standpoint of you focus on that, you execute, you deliver it. Purpose build, what we want to do is create the architecture that allows us to use. We don't want to create huge amounts of SOCs. We want to create derivatives off of a base. If there's a 50 mm chip, 75 mm , 100 mm , 125 mm, use those. What you do in the architecting of the space is maybe there's a section where there's IP block. It's almost very much like what we do in the game console. The amount of the game console chip that we create, the vast majority of it is a standard IP, standard AMD capability and IP advantage.
What we do is we wrap a certain set of segments with specific solutions that create. What we want to do with Lisa is drive those adjacent spaces, those towers that give us that lift in terms of TAM and return. We don't want to go spread this thing. I'm a volume guy. I've always been a volume guy. I was a volume guy where I came from. Volume's your friend. It's a good thing, especially if you scale it on expense and you can execute the heck out of it. It's a good thing. We've got to expand because this market is beginning to change. We think that's how it opens up. Think about it one step further.
What if you get a large player that plays across tablets, maybe even into smartphones, across laptops, desktops, into smart televisions, into appliances, and you create an architecture using our chip solution with some of their specific IP that allows them to knit that solution together in this converged stage. That might change the way people buy processors in the future. It would help them differentiate. That's in the future, but we need to begin that work now. Fix the execution engine, get the scale, make sure you deliver on the cost model and scale that cost. Make sure that if you can create that different model, if it's truly changing, which we all believe it is, that's the opportunity. Don't spread your supply lines too thin and get yourself spread across a wide base. Remember, the way we architected the organization, Mark drives all of the technology.
He's driving across technology and engineering with his team. Where we make some investments into some customer-facing and support levels to open up some of those spaces is in Lisa's business around the go-to-market, and then around the sales coverage should be consistent across regions. That's the one area where you'd invest and you wouldn't spread it across. You want that synergy across that common base. That's the engine that creates the innovation. I hope it helps.
Second to last question.
One comment here, cannot let this go. Of course, there will be a good business case around every decision.
We will not resort to it here.
I do want it for fun. It was important that we laid the groundwork, especially on the supply chain side, in order to prepare ourselves for that. I think that is a much bigger challenge, managing your supply chain through such an approach. We've done good work. The rest is a matter of business cases and trade-offs. How we look at the space today, we think there is significant volume around customized products around specific workloads or around specific customers that make a lot of sense for this approach.
Just from troubled gardener, as we're looking at moving towards the SoC, obviously in the consumer products, you have those IPs. Within the server world, HPC, you have those IPs. Could you give us some of the colors around the other set of IPs you have in the other markets? Given the market natures, each market is very different from the IP requirement. Many IP owners today are actually acquiring the processor cores to build their own SoC. How would you compete in those markets?
Yeah, it's a good question. First of all, it's about unlocking the value of our IP. As I showed you today, across our CPU, across our GPU, rich multimedia, rich interface technology, very high-speed interface technology we've been leveraging in our graphics products. It's been key to the type of leadership that we've been providing. When you say, how do you compete? Other people are going to a vertical integration. Every time you see one of those vertical integrations, what you're seeing is others getting locked out from being able to play. There's a very rich ecosystem out there of IP. We're going to continue developing, differentiating IP. We're going to look and have enabled with our approach to be able to bring in what I call IP that's complementary. We see a rich opportunity to bring that in as appropriate.
The third source is actually where, in unique cases, there's a deep partnership with a customer that actually has their IP that will fold into our ecosystem going forward. There is a combination approach. It's going to be a very dynamic world out there. We're at a great starting point because we have a treasure trove of IP and we see great complementary IP out there that we can add to it.
Lisa, you've got a lot of background in this.
Yeah, I think Mark covered it well. The IP that we have is tremendous. I think there are also a lot of people who are looking to partner and really marry that IP with the strong ecosystem that we have. I think we see a lot of opportunities as we go into these targeted verticals, not just to go by ourselves, but with partners and with other customers, with our customers.
Thank you.
Great. Last question here in the middle.
Thank you. Question on two questions. First on competitive landscape. You mentioned user experience many times as a differentiating factor. Sure. One of your big competitors is soon going to have 22 nm products. I know that by itself is not differentiation, but they will also have DX11 graphics support, which was missing before. Rory, from the background that you have, looking at AMD from outside, when somebody comes to you with 32 nm products with certain experience, and somebody comes to you with 22 nm products with certain experience, what are the decision criteria that you go through in selecting what you will use at the high end?
Sure. You know, a business unit in an OEM is going to take the time to really go through the analysis of the experience. They're going to break it down in terms of the workload, how does it connect, how does it perform in terms of graphics activities, what's the product targeted for? Based on that, they'll do a technical assessment that matches up. We do well in that. That's why you're seeing the strong design wins. They take and boil that down in terms of the market opportunity. They look at the landscape across their set of portfolios. I was in a global OEM, so we looked across 12 regions. We'd see how that was going to be positioned. We saw where we could put it in terms of pricing, what average pricing could we put it at.
We'd look at it compared to competition, and we'd run through to see what the return on it. Again, remember, there's only five parts in the unit itself that really drive the whole cost. I mean, the way you make your choices on those five parts, you know, hard drive, memory, processor, operating systems, and glass, those are the five things that really matter. If you look at that competition, run it through the numbers, and you look for the return, because you're running on a couple few points of return, you want to get the scale, you want to beat competition with a differentiated solution that gives you a return. That's how we would break it down. I think we look very well in terms of the capability.
When you add the value proposition in, it's a really powerful answer, assuming you execute with precision and you can deliver on the commitments. If I'm COO of a big OEM, I want to make sure that I'm going to deliver the volume in the quarter. I'm not going to risk, you know, 3% or 4% of volume if a company can't really execute. That's why our focus on execution, that's job one. Every day meeting that commitment opens up so much better opportunity. Hope that helps.
Yeah. Just one follow-up, if I may, for Thomas. Trajectory of gross margins and how we should think about that this year. I think you kept the same 44%- 48% band, but you have a few events coming up. You have the wafer price negotiation, and then you will also have more 28 nm products later on through the year. Some positive, some negative vectors. Just conceptually, how should we think about gross margins?
How it's going to develop over the year? I think that's a very fair question. I tried to give you some color on it in terms of the headwind and the tailwind we have. I think it will be a development that's going to pick up towards the second half. We started off well with the guidance we gave for this quarter, and we take it from there. We should accelerate as we move through the year.
Great. That concludes the question and answer session. We're going to wrap up now to hear from Rory if the other speakers would like to leave the stage. We'll just move the chairs.
Okay, let's move to the closing slide. I think what's interesting, and I think what we saw today as we went through it, I thought you get a sense of how the organization and team is coming together. Here's a set of leaders. First with Thomas, taking us through what we've started and how we've continued to execute and deliver on it. Those numbers, the financial improvements, the way we, you can go to the next slide, the way we've moved to the commitment, the way we've built that base around the financials, it's strong, it's consistent, and we're continuing to move forward. As he mentioned, we're already being identified in terms of a credit rating, watch for improvement. That base, that sound foundation is something we have to continue to finish what we started. We heard from Lisa, and we see this focus on the market.
We move away from an environment where we've been only focused on an incumbent or an internally focused technology race. It's not about that moving forward. There's a lot of processing capability already there. It's the experience that we can create and how we can open up this market to the four and a half billion people across the planet that aren't connected today. Lisa talked about a roadmap that delivers this year and next, but also opens up the opportunity to tackle the technology inflection points that are occurring in front of us. You hear from Mark, and Mark talks about how we move to an architecture that lowers our costs, because I'm definitely a scale guy. I like volume. It's our friend. We need an architecture and a solution that allows us to deliver it consistently every day.
Our job is around execution, doing what we say and owning what we do. It is about ownership and a commitment. When you saw the team here today, I hope you got the sense of the alignment, the united kind of thinking. It's not just one person's point of view. It's a synergy across the organization. We have some of the best and brightest capability in this company. They see these inflection points. We're trying to unlock the capability of this company to move forward and differentiate our capability to deliver value to our customers. That's what it's about. It's about low power. It's about video and graphics. It's about that next 2 billion customers. We're going to innovate. That's how we create sustainability, a company built to last. We've created a foundation with our focus on the financials, and we'll finish what we started there.
We'll also create a culture that unites every AMDer across the planet in a singular focus and sees this opportunity that's in front of us. When we align that team and we create an organization that is committed and passionate across every commitment, every customer, every interaction of every single day, every commitment matters to this company. Every one of us owns that commitment. We've got to live and breathe it. We've got to drive that. If we capture that kind of drive, passion, and energy, and think about AMD, it's got that entrepreneurial spirit. It's got that can-do. It wants to fight that battle. If this inflection point begins to emerge, as we are suggesting, which it's going to, and the proprietary control points that have been in the space begin to break down, we are uniquely positioned to take advantage of this. This is our opportunity.
It is our time in AMD to step forward and capture this inflection point. We've seen them before, and we've missed them. This is our opportunity right here, right now, to begin to change our execution, to become an execution model that delivers and be able to deliver on every customer commitment, and then to use the innovation, the creativity of the minds that we have in this team to unlock this future. This is our time. This is a different AMD, and this will be a different outcome as we move forward and write a new future for this company. Every one of our 10,000+ employees across the planet are coming together around this concept of this opportunity, first starting around execution, build on our innovation, and capture this inflection point. I think we're uniquely positioned to do this, and I want to finish the way I started.
I want to thank you for taking time out of your day to come here and spend with us in person and across the internet. This is a different company and a different time, and the industry is changing, and we embrace that change, and we want to facilitate that change. I hope you enjoyed the discussions we had today and the beginning of the conversations we'll have over the coming months and years as we embark on this journey together. I appreciate the time, and I hope you enjoyed this time together with this AMD team. Remember, it is our time to seize this moment today. Thank you.
At a point that we're about to get the revenue. We've been really pretty focused this last year trying to deliver that. I'm going to talk about the fabless strategy today. I'm going to talk about our overall manufacturing capabilities. I'm going to talk about how we fit with the technology teams and drive forward into that. You heard earlier, it's no longer about speeds and feeds. Very important. Where does it start? It starts with our product roadmap. What does our product roadmap do? We need to understand what technology it takes to drive that. Then, of course, we get back to our overall manufacturing strategy. Manufacturing partners, and then, of course, that leads to foundry partners. Foundry partners, of course, are only part of that, and the foundry partner becomes the lobsters pulling the tent.
Everybody understands the cost of wafer fabs these days and the importance of wafer fabs, the capabilities that they bring with their technologies, and so on. A very important aspect of this for us. I'm going to talk about the semiconductor supply chain in general. A very generic view of the semiconductor supply chain, broken down into three major components. The first part is the part where we get engaged, the start of the partnership. I mean by partnership, there's a difference between supplier and partner. Partners have skin in the game. Suppliers only supply products. It's very different. I'm very serious when we talk about partners. The foundry or manufacturing partner selection process, the fab process itself, and we'll break it down into fab front end of line and assembly test back end of line. Bump and sort is kind of fungible.
Those that know about the manufacturing flows know that the bump and sort can be a part of the fab side of life, or it can be a part of the assembly test partner side of the house. Probably we should stop here just to think about fabless. Fabless works for AMD. I'm going to repeat what you've heard again this morning. 30 million Brazos, 10 million Llano, albeit it was late in the cycle, but it was 10 million that we produced last year. The world's leading GPU. Fabless works. Fabless AMD also works for our fabless partners. What do AMD give our foundry or manufacturing partners? We give them access to world-class R&D resources. We've actually got people stationed in every major supplier's location. We partner on a daily, weekly basis with all of the key partners we've got.
On top of that, we've got access to the leading device analysis labs anywhere on the planet. We've got phenomenal test chip capabilities that we design and run through as a snowplow to fast learning for our foundry partners and technologies. It's a two-way partnership. We get and they get. By the way, it's not always rosy. Those of you that think about this, it's not a date, it's a marriage. Anybody that understands that knows it's not always rosy. Talk about the foundry selection criteria. For that, you could think about any supplier selection criteria. For me, there's five critical elements. It starts with technology capability. Does the foundry supplier have that capability today, or will it have it in its roadmap such that it will meet our demands, such that it will meet our needs? If it doesn't have it, will it have it?
That's the point in time ahead of the process where our R&D teams get together and start working very, very intimately with our potential supply base. Then you've got the next three aspects of this, which are commercial, performance, and flexibility. It's no use having the technology if you can't get the price. It's no use having the price if you can't get the factory performance. It's no use having the factory performance if you can't get the capacity or the flexibility. Commercial, you want commercial, the best commercial arrangements over the life of that product or that technology. Obviously you want the best price, but it doesn't always mean that. A lot of leverage you get in this relationship comes at a yield or time-to-market. You really got to structure your commercial agreements accordingly. That's pretty important in all of this.
Again, factory performance, it's no use having the best price and the best technology if you don't have consistent yield, consistent output. Flexibility, it's no use having all of that if you don't have the capacity to deliver, the flexibility to move your schedules. Many of you guys are involved in forecasting. One thing that we all know about forecasting, they'll never be right. It's just a question of the degree how wrong they'll be. It's no different for us. Our forecasting, as much as we try to make them as accurate as possible, they'll never be there. What we work with are the right foundry partners to create that flexibility in our strategy and the ability to move in new products, the ability to do revisions on cycles. All of that comes together to make sure that we produce the products at the right time. All of that's important.
We come to the relationship, the one that surrounds it all. You've heard today about the deep relationships we've got with our suppliers. They are deep, and they're not always pleasant. Everyday calls for a long period of time, intimate, sitting down, often frequent, everyday drumbeats. We're on calls, our team are on calls, or involved in hands-on face-to-face meetings with our supply base every day, seven days a week. That's what it takes. They're not, again, I repeat, they're not always pleasant meetings. Sometimes they are. In fact, let me assure you, they've got more pleasant in the last quarter than they were the previous quarter. You need to work that one out. Part of that partnership is risk sharing. I said earlier, skin in the game. We share long-range planning. We share technology roadmaps. We've got skin in the game. We've got people involved. We've got IP involved.
We've got risk sharing. We have to lean into certain things to meet milestones. We share the risk. We share the dollar cost. Many, many aspects of that in order to wrap around this whole thing. You can call this foundry selection, but it's the same type of selection criteria for most of our suppliers today. Foundry, of course, is a long pull in the tent because of the time ahead that we've got to engage with our technology teams. They're not engaged just at the start. The relationship and partnership goes all the way through till we get the product there, till it's debugged, till it's working, till it's functional, till it's in the platform, till it's actually making revenue for AMD and our customers. This is a very important aspect of what we do. The front end, the easy part. That was a joke.
It's the Scottish humor, I'm sorry. Technology gets more complex as a mass layer increase. Very obvious. Many of you guys know about wafer fabs. You'll know that as technology increases or gets tighter, we introduce more layers. Introducing more layers means more complexity in lithography. Anybody want to guess what the latest immersion lithography tools plus tracks cost today? I'm guessing, but I know it's in the region of about $50 million per tool. Start to multiply that as you look across the wafer fab flow from the start to the end, front end of the line, middle of the line, back end of the line. Every part of that fab is governed by lithography. By the way, lithography is just camera work. It's just cameras. Extremely expensive cameras, but they're cameras nonetheless. That's what this drives.
Most fabs today are designed to have the lithography tools as their bottlenecks because you really want to keep them feeding the rest of the fab and keeping them fully occupied all the way through the process. As you go through the process, front end of the line, middle of the line, back end of the line, you actually start integrating the process flow. You develop the characteristics for the device itself, for the SoC. You've got the high-k metal gate at the front end. You create a lot of the performance characteristics for the middle of the line. Then you've got the metalization process in very dense pitch in GPU in the back end of the process. All of that has to be integrated. That's the challenge in the wafer fab. Wafer fabs today have thousands of processes.
Wafer fabs have evolved to have really, really sophisticated automation to the point where the wafers are actually flying in overhead gantries. It's almost like air traffic control, how you move these pods from one part of the fab to the other. It really is an incredibly sophisticated activity. Now, when you think about the cycle times, think about the cycle time from the front end to the back end. You can see by the rough calculations there, it can be anything from 11 weeks to 13 weeks. Lower technologies are shorter and more complex technologies. It could even be longer than that. When you think about that, can you imagine if you have a problem in the front end of the line? You don't find it out till 13 weeks later. You can think about the damage you do in that fab on all of that inventory.
You can have three months inventory. Guess what? No surprise, the fabs have developed and evolved very sophisticated capabilities to track in line, to understand what's going on. Literally, thousands of SPC charts, tremendously, and by the way, they're automated, of course, tremendously sophisticated camera inspection tools where they're inspecting product as they progress through the line. Obviously, at every stage of that line, fabs have got control and balance and checks and balance before processes step to the next phase of that activity. Of course, there are other techniques, short loop learning, test chip methodology we've taught, and all of that with our partners. AMD is involved in it every step of the way. When the product gets out, the debug, the understanding of design characteristics, is the design good? Where can we tweak or move the design to make it more functional?
Where can we start to do different things in here to help this whole flow and help the foundry, but obviously help AMD through this? The back end of the line is more commoditized. The front end of the line is longer in cycle time. The back end typically is two to three weeks, depending on how you want to play it. This is where AMD truly differentiates. What we have in the back end of the line, we have our legacy internal factories where we do assembly and test, coupled with the factories that used to, coupled with the test and assembly partners that came with the ATI acquisition. We blended them. Virtually every product that AMD makes today is dual sourced. We have the capability to run in two sites, and we've got the capability to run over two, at least two test platforms. Incredible flexibility.
Our own major manufacturing sites are in Penang and in China. We've actually reconfigured the furniture to actually integrate them. One of them used to only do assembly. The other one only used to do test. We've moved the furniture around to make them capable of doing assembly and test. What does that give us? That gives us a phenomenal advantage, and it gives us the potential of four-day cycle time versus three weeks. Obviously, we stage inventory. Part of our postponement or part of our inventory management is collaboration with our customers that allows us to manage the inventory in line and at the end of the line so that we can pull for them. Now you can start to see from Die Bank to the customer's dock can be as short as four to five days from the time we've produced it.
Die Bank, we do pull, and for some commodity product, we push all the way to finished goods. You can see the light blue colors there are inventory activity points where we will postpone or we will pull accordingly, depending on the customer, depending on the customer differentiation that we have. We get the product to pack out. You can start to see how this whole thing starts to come together between the fab side of things and the back end of the line here. The other thing I should state is that this is also a tremendously good cost activity for us. Thomas talked earlier about the CapEx that we spend. We have not spent virtually a penny on any test platform in the last two years, and we're very proud of it.
It's just efficiency and productivity and moving furniture around and changing test programs and changing test patterns such that we move them around. The integration is giving us a real differentiating factor here. When we look at this end-to-end, excluding the foundry selection part, which we know can be quite far ahead, the foundry and the bump sort could be either side. It can be associated with the foundry or it can be into the ATMP side. We've got Die Bank. Then we've got assembly test. The whole thing end-to-end can be 16 weeks. As I said earlier, strategic partnerships and strategic collaboration with our customers allow us to put material in play that we can move much, much faster. That's on a fair service level agreement, and it's all part of our collaborative approach.
Characteristics, as Rory said, you've heard that time and time again: execute, execute, execute, meet your commitments. This activity helps us do that. From a customer perspective, make our commitments. Create the flexibility that allows us to accommodate hiccups. We will have hiccups. Anybody that's ever been close to a fab environment or a test and assembly environment knows that something happens, and it will happen. This allows us to mitigate some of that activity, that capability in the back end, the positioning of strategic inventory. By the way, if you've been watching our inventory, you know our inventory is in very good shape. We manage that accordingly. Internal, we've got to manage our own capabilities. We manage what we're doing in terms of managing inventory. We've got processes that are, we've got consistent processes that allow flexibility. That's the key there.
The last one, of course, is back to the drumbeat. We don't want to have drumbeat meetings with our manufacturing suppliers. We've got drumbeat meetings with our customers. We've got drumbeat meetings internal that we talk about different customer needs and where our pull points are to stay ahead of that. We've got the right set of metrics, and we're focused on them day by day, minute by minute, and driving this stuff very, very hard. I'm at the end of this, and it's probably a good time to get some Q&A at the end of it. We've got a balanced strategy. You've heard that we're partnering with the giants of the industry in terms of foundry. We also partner with the giants of the industry in terms of the back end. Our material suppliers are also the biggest guys out there. We've got relevance with our supply base.
It matters a lot when you want to accelerate things and move things faster or create flexibility, push, pull, faster lots, new product introduction. At the end of the day, we've got to manage this to meet our strategic intent, making sure that we meet our customers' needs. Bottom line is, for me, execute, execute, execute. You've heard that time and time again. With that, I will stop. Are we going to do a Q&A now? I'm happy to take questions, yeah. Sure.
Just when I'm realizing this is your gospel medium in the context of technology, availability of deals and stuff, I'm obviously connected to that. Can you talk a little bit about your experience?
Yeah, we're delighted with our Radeon experience. Straight out of the shoot, our products come out. How do I feel about 28 nm? I'll try and paraphrase, but I hope I get it right here. How do I feel about our 28 nanometer experience on Radeon, so graphics products? My response to that is that I feel very good. We've come out of the shoot running. Our yields are bang on where we predicted them to be, and we're supplying product to the market today. Okay. Thank you.
You mentioned hiccups entering your talk. During the fourth quarter, AMD alluded to a problem with 45 nm yields that affected some of your deliveries. Can you give us a little bit of color as to what happened there, how you corrected it, and why it's not going to happen again?
Yeah. There were two things. There was a huge demand on 45 nm. Normally, we would have had inventory, but the demand was very significant very quickly for the channel business. What happened was there was a misprocessing event in our foundry partners' space. The good news is that's been corrected. The good news is that we've recovered fully and that we're running back to normal. Because we've now got a better view on the forecast, we'll actually build some strategic inventory there to mitigate that in the future. That was an unprecedented demand that spiked on an older technology, which surprised us.
Explain, or just shed a little bit of light as to how they screwed up on the 45 nm?
If you guys know wafer fabs, you know that there's multiple opportunities to screw up. Multiple. Every day, there's something new. It was a misprocessing event. I'd rather not talk about that, but it was a misprocessing event by our partner. We don't like them. We force, you know, our quality teams are all over it. We force ADs to be driven down. We go back to, you know, detect, prevent, avoid, do all of that stuff. The main thing for us is to learn from it. That's, in some ways, to some extent, the cost of learning. It did hurt us a lot in Q4, but there's been a full recovery already.
Can you talk about, as your global foundry business and TSMC all converge to CMOS, what are the rules that the different foundries and their libraries put constraint on your design capability? In other words, are they truly foundry neutral? In other words, as long as high performance, 20 nm, it really doesn't put any design constraint on your APU design teams, how they design the different chips, whether they use for global foundries or TSMC.
I'm not sure I'm really following your question.
Yeah, I was wondering, since the 20 nm between foundries aren't exactly the same, what constraint does it represent for your design team?
Okay. The constraint there is one of time. In fact, I've got my colleague Chekib here, and he can help me answer this. The constraint really is one of time. Where this is part of the partnership selection process, where upfront you really start connecting very quickly, and that's part of phase one. Is the technology available, and when will it be available, and is it what I'm going to design to? We start that process with, and obviously we want to know where the transistor is targeted and so on. We start that process significantly ahead. It's painful because, you know, the standards are different, as you say. It would be great if, you know, and I know in older technologies, people consider the technology to be T-light, and that's become an industry norm. In advanced technologies, that's just not possible.
We've got to structure our design resources accordingly.
Yeah, I think what I would add, every single supplier has had their own specific process technology and library. We cannot really have one unique library where we can develop all our APU and our design, which will work naturally in all these different foundries. That's the nature of optimizing the design with these specific libraries. We don't call them constraint. We call them like input as the process technology, ground rules, and libraries we're getting from these different suppliers.
Thank you, Chekib.
Yeah, just to dovetail on that question a bit, more along the lines of the assembly and test, and correct me if I'm wrong, you said that you guys had not made any spending on assembly and test equipment for the backend support. Was that correct?
I said no test or platform.
No test or platform.
Where we do specific SLT activity, yes, we put a small amount of CapEx in for that, but it's very customer specific.
As you go forward in levering IP libraries from your foundry partners, do you run into any issues with design for test limitations, and do you see a potential for having to increase or to add some spend to accommodate for your foundry partners maybe doing design for test on more advanced equipment than what you guys are currently deploying on the backend?
On the contrary, I actually think that design for test will help us with CapEx in the backend. I actually think what you're asking for is what I'd love to see from a supply perspective. We actually put more features into our designs itself. I actually think that would work to the advantage of what my CapEx spend is in the backend. Of course, it's everybody's intent in the backend to push the advantage of what my CapEx spend is in the backend. Of course, it's everybody's intent in the backend to push the failure. Product is going to fail after you produce it, you know, in the wafer form. We'd rather we cut it at wafer form and sort than package it and have it fail there. You're right.
The more DFT features you can put into that design for manufacturing, get all of the stuff cut at sort, it just means all of the downstream activity gets less. What AMD doesn't do today is we don't sort our own material. We leave that to either a combination of the foundry in some cases or the backend test and assembly guys in other cases. Our product in that respect is very fungible.
Do you guys, forgive me, do you guys purchase wafers, or do you guys purchase finished?
We purchase finished wafers. Yes, for sure. In fact, it's different models. It's different models for different ways of doing it. Right now, of course, you have people talking about last year's WSA was actually finished goods from Global Foundries. A lot of suppliers, it's actually we buy wafers that are expected to yield a certain percentage. That's a question at the back.
John, can you talk a little bit about Global Foundries building a state-of-the-art facility in upstate New York and sort of your plan and how that's going to affect our supply chain?
I really can't. I've not had the opportunity to go there yet. Global Foundries is a long-term partner of AMD. We will be producing some product there at some time in the future. As to what, we haven't yet sat down with them and discussed that. It's part of our ongoing discussion, as you've heard earlier in today's discussions. It's inevitable we will be in that multi-factory at some stage. We're obviously keeping very close tabs on what they're up to. We'll watch that closely. We will be in there sometime in the not-too-distant future.
Yeah, this morning they mentioned that they're going to have a BGA version of Trinity. I'm wondering what impact, if any, that will have on your backend operations.
Actually, not a lot. We're already prepared for that. Even to the point where we go different package Z-height. We already understand there's three different packages coming out in Trinity. Our teams are prepared for all of them, even.
To the finished package we can produce, we're actually very excited by that. The potential of VM, these form factors are really, really strong for us, and our package partners in this case are in Japan, and we're working very closely with them as well.
This might be a little bit out of what you normally deal with, but obviously, your OEM customers have to take the BGA chips and integrate them into their manufacturing flows much earlier in the production process than they do for socketed products. How are they dealing with that?
I don't know the answer to that at this point, but I do know that we've created samples already that are in our customers' hands. We could follow up on that question for you, but I don't know the specifics at this point.
Hi, can you talk about what your plans are to get to the 22 nm node? I guess for this year, whether there's been some talk on the blogs about 28 nm yields at TSMC not being so great, does that push the plans to 22 by a certain degree?
I think you have Mark Papermaster talk about that earlier on, that we really are looking to, you know, as part of our strategy. I'm going to let Chekib talk about that in a second. Let me just reset that, you know, to the question that was asked earlier. I'm very happy with my 28 nm GPU yield. We're supplying product to the marketplace. Sometimes blogs are accurate, sometimes they're not.
I can add a little bit about it. After 28, by the way, we are looking to 20. We're not looking to 22. That's the new kind of step nodes. Between 28 and 22, there is not enough hyper node for us to justify that move to that 22 nm node. 28 nanometer as a node actually is a very good node in terms of performance, power. The challenge we have in 20 is not what's happening in 28. It's the 20 nanometer per se in terms of cost and in terms of performance and density it's going to provide us. It does require double patterning, as you know. That's costly. We are really very carefully looking at what kind of product can go to 20 and how we can leverage and take advantage of 20 to really synchronize ourselves the right timing to jump on 20.
Two questions. One, just on your GPUs at 28, is that a high-k process? The second question, can you tell us how long, how should we be thinking about CapEx spend over the next few years? $200 million the last few years seems like a lot for back end for a fabless company. Can you walk us through? Does it stay at those levels? Why is CapEx so high and not lower?
Sorry, I can't see you speaking here, but oh, sorry. Yeah. Our CapEx is total company CapEx. It's not just hard. It's including IT and other things. Thomas can tell you more about the total total spend. That's not back end. Back end CapEx is very low. It's maybe in total to bring up new packages. I'm going to give you a number that won't be right, but it's in the region, I think, of about last year, about $60 million. A lot of that came from customized, what we'd call SLP builds. We build the hardware to do secondary testing in our back ends. We also take the hardware to bring in new packages. That's where that comes from. The $200 million build didn't come from the manufacturing side. There are other aspects of that. You know, I'm not going to quote our suppliers' technologies. Yeah, sorry.
Thanks. The question kind of comes back to a bunch of questions, I guess, in that it takes a long time for you guys to partner up with your various suppliers. Clearly, TSMC is different from GlobalFoundries on a number of levels and a number of ways.
Sure.
To the degree that one partner isn't able to meet your needs, call it GlobalFoundries not being able to make it below 32 nanometers, for instance, and you guys need to make that move and you need to make a shift away from GlobalFoundries. How long in advance of that transition do you need to engage with TSMC to transition your designs that were originally architected at GlobalFoundries to make that move over to TSMC or vice versa, for that matter? It sounds, from what you're describing, it's a very involved partnership and it's very specific in the architecture. It seems that it's maybe not as simple as many of the blogs have made it out to be that somehow later this year we're going to have, you know, the follow-ons to Interlagos coming out of TSMC.
Yeah, like I said, don't believe all the blogs. There is, in the front end of the process, the design gestation period where you're looking at the targets, the simulation targets that the foundry creates that you want to match to. That can be anything, you know, and again, Chekib is the expert on the design side, not me, but I think it's anywhere around about the year timeframe. That's something that you really don't want to do because changing is expensive. At that point, you've committed an awful lot of effort. That's why the partnership up front is so critical to really understand where you're headed. When I said it's not just about the technology, it's about all the downstream capability, the speed of ramp, the capability of NPI, the actual involvement and drive of the test chip methodology.
You get test chip methodologies of snowplow, so you can actually see what we're doing. In fact, we've engaged that in both our foundry partners today. The test chip methodology is actually very much appreciated by our foundries. I don't know if you want to add some commentary to that, Chekib, about the time up front.
Can I add that at least for the...
Can you hear me?
For the development part, for the design part, it's roughly between 9 months- 15 months. The reason why it's a range is because it depends on the design style. If it's fully synthesizable, we just kind of drop the new library. That can go very quickly. That's 9 months. That's 9 months, by the way, just the design side. There's another part, which John is referring to in terms of getting the test site, getting the manufacturing ready. It could be up to 15 months when you talk about a very highly customized unit microprocessor, for example. That's really where the range comes from. It takes a while. These are not simple things you can just decide tomorrow I'm going to go to another foundry. It's a deep partnership. It's a complicated process technology. It's a very heavy system on chip kind of product.
It's definitely a lot of effort and a lot of work.
Moving below 2 W or to the 2 W and below power envelope, what process technology do you guys feel is going to take for you to get there? What percentage of the reduction in power is going to be from the process technology versus from what you're doing on the architecture side?
You're stepping outside of my area of expertise. I'm going to ask Chekib to join me up here. That's going to make a lot more sense. We'll try and answer your questions nonetheless.
Yeah, I think below 2 W, you have to be a little bit more precise in terms of also, again, any other product. If you refer to where we are heading in the tablet, going after 2 W, we're definitely going to need much more appropriate process technology specifically to control the leakage. It means we will be having transistor with higher VTs. In any kind of process node, let's take 32 or 28, there is a lot of variance in terms of different process technologies in the same node, one which will allow you to have a very high performance. You will accept some more leakage in terms of the transistor, but it gets you much higher performance. When you go to the very, very low power, you really have to have much less leaky transistors and you have a different flavor of that same technology on that node.
In order to get the 2 W , they call it the low power technology, LP or LPE, depends on what kind of really suppliers you are talking to. I'm sure you're on me here.
No, no, you stay there. You're doing well. There's one more question down here.
Hi, guys. I'm John. Can you talk about the, you mentioned the inventory at various stages of manufacturing? Can you talk about...
You talked about inventory staging at various manufacturing stages. Can you talk about how you think of that in terms of percentage of sort of run rate business and how much you want to put in each bucket or how much you're comfortable with putting into bucket and how that may change as total demand goes up or down and pigeonhole around that?
Yeah, that's a... Let's just take a step back from that. Think about product life cycles. The question we were asked on 45 earlier on. You've really got three cycles: the ramp cycle, in which case you might be in a risk situation trying to understand really the pull on your product. You want to, and you know it's good, so you'll build more and you'll overhang. You've got the static state, which again, you want to say, hey, this is pretty good. I can build inventory for Die bank or finished goods, depending on how commoditized it is. Then you can, at the end of life, where do you really want to go with that? There are three cycles, but when you aggregate it all together, what we're really looking for is a total of six inventory turns a year. That's our healthy operating numbers today.
That's where we want to be. We've got to manage within those three cycles. Obviously, product by product is very different. Customer by customer. We actually have Die bank, which is really post sort. We can pull from that. As I said, for key customers, we can do that now in less than five days. We'll keep product at finished goods, which is post test, which is ready to pull straight to, either it can go into a box or it can go straight to our distribution centers, or it can go to a customer-owned hub. There are a number of different plays that we make by product, by customer. We differentiate service levels for our customers that way also. You see that anywhere you go today, there are differentiated levels of service. Coach, business, first class. We have customers that fall into those categories.
Yes. I'm going to try and re-ask the question on the GPU 28 nm to see whether or not it's high-k or Xeon, because TSMC actually supplies both. The second question has to do with your roadmap this morning laid out 28 nm in 2013 across almost all your products. Will that be, will you lean toward a more high-k , more Xeon, or a mix? Thanks.
Since you asked twice, you're going to get to the answer. All of them are in the high-k middle gate.
Across the board. Over here.
I'm curious in terms of your two foundries right now. How long would it, or do you anticipate a day where the two will be completely interchangeable for all your products? What are the roadblocks for that?
Oh, I don't ever see that happening. I think they're very different platforms, very different setups. You know, the GlobalFoundries is part of the ISTA activity, this GlobalFoundries at Samsung, IBM, STMicro, and a few others. You've got the TSMC platform. I don't see they will ever reach that collaboration. They're competing against each other pretty seriously. It would be wonderful for us in the industry if they would. It would make our lives so much easier, but I don't think that's ever going to happen.
Thank you.
Thank you.
Is that fun?
OMG, Louis, you are killing me here.
That's it! I cannot work with these amateurs any longer. I am a serious actor.
You said allergy season is something?
[Foreign language]
Hello, welcome everyone to the client and graphics breakout session here at the 2012 Financial Analyst Day. Big welcome to those of you here in the audience at AMD. Also, a great welcome and thank you for joining to those of you coming via webcast. I'm John Taylor. I lead product marketing for AMD, working in Lisa Su's global business units. I have a cohort on stage with me as we present to you demo theater here for the next hour.
Matt, Manager for the GPU business unit at AMD. As John says, this is an action-packed session. I think we got nine demos, which may be a record for a short session like this. We're looking forward to that. First thing I did is I want to talk a little bit about the demo you saw as you came in. The most interesting thing about it is that's not a video. That's a demo. That's a demo that's rendered in real time, driven by the world's fastest GPU, the Radeon HD 7970. It's on Eyefinity, so it had three driving three displays there. What's interesting about it is that was created by our in-house demo and research team. Maybe, I don't know, a few years ago, each frame of that would take minutes to render. Now we're doing this in real time.
It's an example of a type of experience that we want to deliver to every screen with our graphics technology. If we step into the demo mode, we'll take a little bit of a closer look and look at a couple of other cool things about the demo. As I said, it's real time. You can walk around in the demo. We're just moving around through the demo. You went through the wall there. That's what I think.
Essentially, Matt, we're looking at objects, textures, characters that are as complex as what's used to make a great CG film today, but rendering it instantaneously.
Yeah, absolutely. You can also start the show and walk around while the demo is running, which is pretty cool. It's a very complex scene. If you put on wireframe mode, you'll see it's running about 60 million polygons a second. Every pixel has about 30 light sources being rendered on it at one time. Take off the wireframe mode. Lighting is very important in this. If you notice, our hero here, Leo, is setting up a shot, the perfect shot with his puppets. He's painstakingly made all these puppets and the clouds and the castle. Lighting is very, very important. This demo uses compute to calculate those lights. If we turn on the lighting mode, these are all the lights in the scene. In this scene, we can do up to 2,000 separate lights. It seems obviously very complicated. A lot of compute going on, and then it renders.
It's using the GPU for compute as well as rendering. You also get bounce lights. One of those light beams can hit a reflective surface, for example, and it will bounce and create some global illumination. Let's take off the light. Another effect that's in here is called depth of field. Depth of field is a technique that's used in animation whereby you draw attention to something, draw focus on a character or something, and you do that by blurring the background or blurring other images in the scene. We can turn on depth of field. Oh, it was on. Now you turned it off. Turn it on, and you'll see that the dragon in the background blurs out. Fairly complicated, yet subtle, but makes it more immersive, makes it more of a better experience. Now, final, go ahead.
I was just going to say, when Rory talked earlier today about it, it's about that end user experience. You can think about this as relevant to the next generation of content generation capabilities and what the next generation of interactivity in games, the kinds of characters and things that you'll be able to manipulate in that kind of a world.
Yeah, the other thing that's in here is very complex, very complex materials. If we zoom in on, for example, this cloud here, it's made of, Leo's made that, hand stitched it, you can see, out of silk. You can see the sort of the shimmer off the fabric. Take a look at the grass made of felt and sort of a canvas. Where's he going now? Oh, you're looking at the background there. Zoom in here on the knight. The knight's very shiny, and if we get at the right angle, we'll see the rest of the scene reflected in his armor. You can see it there on his helmet, on the shield there. What's happening is it's being reflected. The rest of the scene's being reflected in there. It's very, very complex stuff and very, a little bit entertaining too, I hope. Thanks, Jason.
Okay, so this being demo theater, we could not resist starting out with a demo. We'll jump in now and give you guys a sense of exactly what Matt and I will cover over the next 45, 50 minutes or so. We'll pick up some of the themes from earlier in the day, but try to bring them to life for you in a meaningful demo form and a few additional data points that you perhaps didn't see in this morning's program with the executive staff. We'll start out talking about those major trends in the industry that you heard Lisa Su speaking to. We'll showcase how our technology leadership can be parlayed into opportunity and differentiation for AMD against those major trends that are developing.
We'll really showcase more and more, again, Rory, as he hammered the theme about execution, he hammered the theme that it's about what the end user experience is. We'll give you a sense of those next generation of experiences that we're enabling. Finally, we'll close with something that really serves as an on-ramp, if you will, when we talk about accelerated applications, show you a few of those. The third breakout session and final breakout session today with Phil Rogers and Manju Hegde will really deep dive into how we will continue to generate new generations of accelerated applications through our GPU and APU architectures going forward. This should look familiar. This is what Lisa spoke to earlier. In general, we have some demonstrations for each of these major trends. Let's start with natural user interface.
Some of you might be questioning, what does natural UI have to do with all this work that AMD has been putting into parallel processing and APUs? You need to think about something. Let's take something familiar like the Xbox 360 Kinect technology that's now having a lot of interesting activity with developers happening around it. That process of a Kinect with a depth camera and a traditional type of a web or HD camera looking out into a room, if you've ever looked at some of those videos of how those are hacked to see what the world looks like to a computer through the eyes of an Xbox 360 Kinect, it's thousands and thousands of points of light.
That act of processing what's changed in this field, where are the faces, where's the arm, how far is it from me, how far is the person from me, perhaps I should change the way my user interface is represented, whether they're 10 feet away or 2 feet away, all of that lends itself to acceleration through parallel processing. Similarly, if you think about even something like all of our multi-touch devices that we use now, and think about something like a 4, 5, or 8 megapixel photo, and you want to manipulate it with multi-touch, having a true graphics capability represented in even those very thin types of form factors gives you a better overall, more responsive multi-touch UI.
Another trend is moving to higher definition, moving to more pixels on the screen and more displays. You've seen that in consumer TVs going quickly from standard definition to 720p, 1080i, 1080p. You're seeing this also from a PC point of view. If you look at displays maybe four years ago, you get a 22-inch 16 by 10 monitor, would be about maybe probably $380 or so at that time frame. Now you get a 24-inch 19 by 10 monitor for about half that price, 20% more pixels, half the price. Driving more pixels. We're seeing this trend continue. The other thing is from a multi-monitor display point of view, one of the things that we think has driven that a little bit is our own technology called Eyefinity. Having more displays for productivity and for their activities is becoming more and more important.
If you look even practically, if you buy a large display, like a 30-inch display, that's still over $1,000. You can get a 3 by 1 setup for about half that with the 19 by 10 monitors.
The third, working from the top row here, social games. Of course, it was a big year for gaming in 2011. We expected again in 2012, we had the world's most successful in terms of sales game ever when Battlefield 3 was introduced. That's an AMD Gaming Evolved partner. Also, so many of the headlines went to the developers of these new kinds of social games. Think about the kinds of games where you connect with your friends on sites like Facebook or other social networks. We'll delve into a little bit more here and show you some demos of we've seen really the first act of social games. Those games have been rendered by CPUs and not by GPUs, and therefore, I would argue not rendered to their fullest, not the kind of experience you could get, say, from a game console.
We'll talk a little bit today about how you can get the best of a social experience with these games that so many mainstream users really have gravitated to, but taking it to the next level in terms of the visual experience. When it deals with cloud, we have demos related to that. From a cloud perspective, the things that, for example, an APU or a better graphics capability can deliver in that system. Of course, there are opportunities like compression and decompression. If you think about how much time you spend looking at something like streamed content on a cloud device, a tablet, or a convertible, or a hybrid product, with our APU technology, you have the ability to do things like dynamic post-processing of a streamed video feed from a Dailymotion or a Baofeng or a YouTube as you're watching it.
We can differentiate and make that experience even better, even in a very thin client-type cloud form factor.
The final one is collaboration. We all want to work together more closely. There are many examples of this, and it sort of relates to some of these other trends with social games, that sort of thing. If you look at just even video conferencing, I have a friend of mine who lives in Munich, and every Sunday night he has dinner with his parents who live in Switzerland. He does that via video conference. We all travel a lot. What if we could have what's pictured here, a six-way video conference with really great quality, be able to save those air miles, save those travel dollars? That sort of thing is really something that we can take advantage of with our technology.
Definitely. This gives you, I've talked a little bit about where applications today can be accelerated. This is a look using IDC research at the most popular applications that end users are engaging with today. You'll see some very familiar ones at the top that are kind of old school, like email. While we don't have that highlighted as something that can be accelerated by the APU or doing something like your personal finances, you will see that all of these areas that you can accelerate are places where you can really differentiate visually where the end user or the consumer might have problems or frustrations with what the experience is like today, how long it takes them to do certain tasks, or how complex they are. What do you use your computer for? Hardcore gaming, Matt.
Yes.
A little bit of YouTube, but yeah.
Excellent. Man, that's my own heart. I use it for lots of things. Gaming is one of them, but also video editing. My daughter had an assignment at school. She took a bunch of videos with her cell phone, brought it home, and we had to stitch it together. It's amazing how quick she was. She's moving everything around. That experience can get better. When you render the video, it takes longer than I like. I want that to be instant. That is an application that we could accelerate.
Yeah, absolutely. Video is one of the most dominant workloads, whether you're looking at it from the cloud side or the client side going forward. We talked a little bit about the trends. Now we want to talk about technology leadership.
It's important you have these trends. We want to capitalize on these trends. AMD has led from a technology point of view, both in GPU and with CPU technology. From a GPU point of view, we were the first with an AGP graphics card, first with PCI Express. These are major inflection points in the industry. When we lead in those inflection points, we win. In fact, in both those situations, we went to number one market share. From a display point of view, we're the first with DVI, first with DisplayPort 1.2. From a process point of view, first to 90 nm, first to 80 nm, first to 65, 55, 40, and now 28 nanometer with our 7970.
Absolutely. If you look back similarly from the CPU perspective, you can talk about AMD having the world's first 1 gigahertz CPU, the first to extend the x86 instruction set to 64 bits, then extend it to multi-core, adding HyperTransport for much faster I/O, and then things like integrating the memory controller for much faster speed, lower latency in accessing memory.
Speaking of memory, from a GPU point of view, first with DDR2 on a graphics card, first with DDR3, first with GDDR4, and first with GDDR5 on a graphics card. It's important to lead these inflection points. When we lead, we win. We're first to DX9 and first to DX11. We ship more than 100 million DX11 products. That's a huge number.
It is a huge number. It's important to know that the APU was a big part of that as well. Matt and I believe in teamwork. About a third of those are represented, as you heard earlier today, with all of our APUs. We wanted that first generation of APUs to support that DirectX 11 capability. In 2012, we rolled out a new generation of DirectX 11 products. We believe in it. It's the finding of the industry standard for the experience today and for compute as well.
We're not stopping there. We're not first DirectX 11. We've just launched the 7970. It's the first graphics card with DirectX 11.1 ready for Windows 8. It's the first 28 nm GPU, and it's the fastest GPU in the world. Performance is important. Take a look at this chart. We're looking at 1.3x the performance, 1.4x, 1.5x, up to 1.6x the performance. This is various games at high resolution. Pretty impressive performance here. It's not all about gaming performance. We're really starting to focus on compute also. Engineers have put a huge focus on compute. With the 7970, we're nearly 4 teraflops of compute power.
Now, picking that up in the core theme from earlier today about how quickly we can leverage that investment and that high-end graphics IP into a differentiated set of capabilities in an APU product that you could take through the full mainstream value segments. This represents how we were able to, if you look at that blue line at the bottom, that's your conventional x86 CPU gigaflops performance. Right out of the gate, Llano as a desktop product, you see something on the order of nearly three times the amount of available compute. With Trinity, we step up another 50% to more than 800 gigaflops. We think that third generation APU, the Kaveri product, is looking at more like the world's first 1 teraflop APU. Again, you'll get, I think, a sense today from us how many developers have already worked to unlock that unique capability in our products.
Manju and Phil will take you through after our session how we're making it easier for even more and more developers to access that level of compute. Speaking of compute, I think we've got a demo here that we might want to switch to.
Yeah, we have a demo.
Let's see what you can do with all those.
Yeah, we have a demo. This is called ComputeMark. This is actually a video running a side by side of, put the demo up, a side by side of a run of ComputeMark. ComputeMark is a DX11 compute benchmark. We have AMD on the left side and our competition on the right side with their fastest GPU. It's our fastest GPU on the left-hand side. You'll see some pretty impressive performance from us. This benchmark has about five tests for fluid ray tracing, et cetera, and it really taxes the GPU. You'll see it go, we'll let it run a little bit, go through a couple of things. These are called Mendel charts, I believe.
While Matt is demonstrating this from a discrete graphics perspective, I'd love to be able to do the same thing in the APU market. Unfortunately, as I understand it, our competitor's APU-like product is not able to run this benchmark. We'll see in 2012, I think they're going to be able to step into the arena and we'll be able to deliver those kinds of comparisons for you later to showcase the total available compute in our Trinity APU versus the competitive products.
Let's go to the benchmark results. There are results there. Our score was 2927, the competition was 1756. I did the math. We're 66% faster than the competition.
Good job, Matt.
Thank you.
You're welcome.
I didn't actually engineer that myself.
No, I meant the math.
Oh, the math. Thank you. It's not all about performance, though.
No, it's not.
It's also about power. We need to, I think you heard Lisa talk this morning, I think you heard Mark talk this morning about performance per watt. When we introduced the 7970, we also looked at our power technology. We introduced a feature called Zero Core Power. What that allows us to do is when the system goes into long idle, it allows us to shut down the core of the GPU, reducing the power at that time to less than 3 W. If you look historically, even a few years ago in 2008, in that situation, our high-end GPU would be drawing 90 W. We took it down significantly in 2009, 2010 into the 20 W range, but a massive drop less than 3 W. The reviews for the 7970 were very, very good, and we were very pleased with them.
Maybe it surprised us a little bit about the sort of the raves we got for the way we handled the power in the Zero Core Power technology.
Definitely. Similarly, from the APU perspective, we're equally focused on power. We're also very excited about what Trinity does for us, especially these low-power Trinity parts with that BGA package that we talked about during the earlier sessions in John Docherty's breakout on supply. We've got a couple of different views here just to underscore what we're looking at from a power perspective with Trinity. We're making the first comparison to our Brazos technology. Brazos is really what put us on the map worldwide in terms of power, right? Brought us to TDPs we had not operated at before as a company and put us in a leadership position in battery life in the form factors where we introduced the Brazos technology, which we expect to continue in 2012.
Trinity, which represents a big step up in performance from that Brazos technology at 17 W, is actually lower total thermal TDP, lower silicon power, lower total system power. Now on the right, we're comparing to Llano, which is actually what Trinity replaces. This is the A-series APU in market today. Here what you're seeing is that when you compare performance per watt, Trinity actually represents a doubling of performance per watt over Llano, stated otherwise. Trinity at 17 W delivers all the performance of Llano at 35 W as a dual-core chip.
Talk about immersive experiences.
Yes.
My definition of immersive experience is something that pulls you in, that draws you into whatever you're watching, whether you're playing a game or watching a video, that you're really immersed in it. The frame of the screen sort of goes away. I think one of the biggest changes from immersion in PC in the last several years has been Eyefinity. It's been a differentiator for us for the last two years, but we didn't stop there. We want to continue to improve that experience and to really draw you in even more to that experience. With the 7970, we introduced a new technology. It's called DDM, or Discrete Digital Multipoint Audio. Kind of a cool name. I call it DDM for short because I can't always remember Discrete Digital Multipoint Audio, but it's a feature that allows us to take the sound and direct it to an individual screen.
Picture in this case here, we have a video conference going on. I think that's the grandparents, the baby, and some other grandkids. The baby could coo. The sound would come out in the middle screen. Grandparents say hi. It comes out in the screen where they are. Makes it a much more natural experience.
I'll just chime in. For those of you who may work for some big company that has these kinds of high-end HD video conferencing solutions, you know that those are what, I think, six-figure price tag type solutions to help you connect multiple campuses or with a certain key, very important customer so that you get that sense of, without spending the money on that plane ticket, that you're in the room, you can read each other's body language better. This is essentially delivering.
You know what?
A version of that.
Yes. Why just talk about it?
That's a great idea.
We could demo it.
In demo theater.
Yes.
Yes.
All right. What we set up here is a five-by-one Eyefinity. This is being driven by a single 7970 graphics card. We have two guests here. The way, you know, Eyefinity is not all about gaming. I'll show you that later because it's pretty cool. Here we're doing more of a productivity thing. You could have a video conference going and have several productivity. You got your email going here and looking at some pictures there. We've got two guests here. On the left-hand side here is Carl Wakeland. Do you want to introduce yourself, Carl?
Hi, how are you doing? I'm Carl Wakeland. I'm a Principal Audio Architect here at AMD.
What does that mean? What did you work on lately?
I worked with the teams who helped define what this feature is that we're calling DDM Audio Now. The idea is to take a very simple concept of immersion and extend it to audio. You know, when you're having a video conference, like you guys are saying, you're sitting at a table and the voices are coming from the faces of the people you're talking to. Why not do that with Eyefinity and have the voices of any of your video conference participants come from the actual screen where they happen to be located?
I'll stop you there.
We're going to make it very simple to set up.
I'll just stop you there for a second. I hope you didn't notice that the sound is coming from this side of the room. That's the graphics card directing the sound to that speaker on the right-hand side.
Not only that, it happens automatically.
Thanks, Carl. Do you want to come out here?
No, I'm good.
On the right-hand side, we have Justin Hensley. Hey, Justin.
Hi, I'm Justin Hensley. I'm a Principal Member of Technical Staff in the Office of the CTO, and I work on graphics research and development.
Justin here worked on that demo you saw earlier. Why don't you tell us a little bit about the Leo demo?
One of the really cool things about the Leo demo is that it allows forward rendering systems to use lots of lights.
Oh, the Leo demo. About the Leo demo.
One of the really cool things about the Leo demo is that it allows forward rendering systems to use lots of lights. What that allows is it allows game developers to leverage the sort of the awesome horsepower that's in our latest GPU to give cinematic quality rendering to game players so they can actually be immersed fully in the game.
Awesome. Note, the sound is now coming from where Justin is speaking. Naturally, you turn your head to look at Justin. If Carl said something, you would turn to look at Carl.
I'm okay, I'll say something.
I would turn and look at Carl, so it's a much more.
We could talk at the same time, and you could turn and listen to whichever one of us you wanted to listen to. Yep.
I don't know what to do now. Anyway, thanks, guys. I really appreciate you helping out. Thank you.
Thank you.
A much more natural experience, and it really draws you in.
Absolutely. As you guys can see for yourselves from the slide here, we've got an announcement that we're sharing with you, which is that our upcoming Trinity platform.
Wait a second. You're saying you can do Eyefinity on Trinity?
I'm about to say just that. That's exactly right, Matt. As we've said throughout the day-to-day, it's about experience. It's about differentiation. iFinity enables that for us. We have our first on-stage demo that we'll need to track from the cameras in the back. It's actually a two-in-one or a bonus demo, if you will, where what we're powering this with, and I'm going to come over here and talk about this setup from this laptop on down. This is actually the same reference design, compiled Trinity Ultra Thin, Trinity BGA product that Lisa held aloft during her presentation earlier. What we're showcasing, of course, is that, yes, it has iFinity capabilities. For the sake of this demo, we're not showing you a single large surface.
We're showing you, similar to the demo that Matt just did, we're using each display, actually, to show you a different piece of content or application. What we're doing here is showing that with a single cable out, so think about it, you know, a mini DisplayPort cable out, you would get the best of both worlds or all worlds, depending on how greedy you are. You could have your Trinity that, you know, is sub three pounds in that form factor, so thin that you slip it in your shoulder bag or backpack, and you're not even sure if it's really there. You've got to double check. When you get back to your home office or you get to work, you plug in one cable.
It's attached to our new docking station solution that we're working with our OEM partners and retail partners right now to bring to market in the second half of 2012.
This is John. It's an example of leveraging two of our technologies. DisplayPort 1.2, AMD was first to market with that. iFinity first to market. We're leveraging them both in this.
Absolutely. Yeah, it's all about how we draw that technology leadership through the rest of the product portfolio. That one cable is giving you power, which is great. We all want that when we need to recharge after a trip. From that cable, we're going out to a docking solution. In this case, what we're showcasing actually on the Trinity laptop is we're doing USB 3.0-based file transfer. It's got native USB 3 support. We're also connected to a peripheral here, which is a Blu-ray optical drive. We're playing back Blu-ray content on the second display. On the third display, we've got a little PowerPoint presentation going or a Microsoft Office application that gives a little more details about the overall capabilities for this setup. You get Eyefinity. With these docking solutions, you get the ability to go one cable, power, and light up an entire office.
We think that's a pretty cool experience for the future.
That's pretty cool.
This just gives you guys a little bit more of the details of how that single cable would come out and what that experience could look like, supporting, in this case, up to four displays in a single large surface. Here, I'm just doing the three on the stage.
We want to continue to improve Eyefinity, as I told you before. Another addition to Eyefinity with the 7970 is Eyefinity 3D. That's Eyefinity, three screens, 3D. There's actually a demo in the demo area down there that you can go and see. I don't have enough glasses to do the demo in here.
Do it all in here.
Otherwise, I would if we did. It looks pretty cool and again, just pulls you into that gaming experience.
Absolutely.
We don't settle for just three screens with Eyefinity . We like five too, and playing Eyefinity on five screens is a really immersive experience. Instead of talking about it, John.
Yes.
Maybe we should show it.
Let's do it.
We are going to start actually with the single screen to get you an idea of, oh, I get to play this. This is sort of a single game experience or a single screen experience. Whoa. Dirt 3 is an excellent game, and that's pretty good. Let's add two more screens. You get the sense of your peripheral vision with the two extra screens, and it's a more immersive experience. The thing to note here is that we added pixels. We didn't just stretch that middle monitor over the three screens.
Right. The conventional view would just be the middle screen. You don't get the rest shrunk down to that display. It's all bonus content that draws you in.
Because we don't sell for three, I want five.
Now I'm going to whip out though. Yeah, I remember I spent some time going around the world last year with a Llano-based laptop, showing many of you in this room how Llano could play this game and what that experience looked like on a single display. I got used to this game from that perspective. I remember the first time I played this in an Eyefinity 3 or an Eyefinity 5 setup, it was a completely different experience where you actually felt more of a sense of impending danger, and your heart rate goes up as you go into those hard corners. You can't get that feeling from a single display. When your peripheral vision is engaged and it feels much more like you're actually operating one of these Dirt 3 vehicles, it completely changes it.
Anybody else want to play? You want to play? I'm going to crash now. What you're seeing up here, when you play this in a 16 or a 19 by 10 monitor, on the screen, you're going to have 10 million pixels, 60 million pixels per second. Pretty impressive, driven from one 7970 graphics card. All right. Thank you. Is that kind of cool? Yeah. We are absolutely playing the game, my friend.
I'm so glad you asked that. That was a nice setup. Yeah, for the next section.
We can go back and give you the controller if you want.
Yeah. Our final section before we do some Q&A will be about accelerated applications. There's still actually a lot of demos packed into this grouping. I don't want to steal any thunder from Phil and Manju and what they'll do in the final session. We did want to bring some of the demos to light. This is an area that AMD, while we were truly, I would say, a pioneer in general purpose GPU computing, going back to our original ATI stream initiatives, 2011 is really the year where that caught fire. Many of you joined us at our Developer Summit. You saw how it was sold out. You saw the announcements from Microsoft, OpenCL partners, tools providers, how there is a whole incredible set of energy and growing ecosystem around leveraging special purpose silicon, leveraging the GPU parallel processing capabilities, all under that banner generally of Heterogeneous System Architecture.
What's the proof? What's the evidence of that momentum? As we introduced Brazos and then Llano in 2011, for instance, I know we have a fair number of the technical review press in the audience. They know, and we met with them as we walked through that technology. We had a handful of things we could showcase them, for example, that dealt with GPU acceleration. By our count overall, we think we've got about a tripling of the amount of applications that we have to showcase our GPU leadership, our APU leadership. You'll see there's a few, if the letters are too small, TBAs on here. That's to be announced. The way these things work is that the software developers have their own roadmaps. You'll see some new applications come out between now and the introduction of Trinity.
With some, we're doing some partnerships to bring new applications out specifically timed to that Trinity launch in the midpoint of the year.
Before you leave that side, I want to make one point. It is very important that we work closely with developers to help bring this kind of content to market. We can have the best hardware in the world. We can have technology leadership. We can have Eyefinity with five screens, but you need content. It is key that we work closely with developers to bring, be it games or other applications, to market to complete that experience.
That content is the experience.
Absolutely.
A chip is not an experience. The content is. All right. Let's, speaking of content, here really we're talking about a platform for all kinds of other developers to create incredible experiences. This is the new Adobe Flash 11 with their Stage 3D capabilities. Remember the key trends we talked about, the idea of social games. I said that most of the social games, and I think something like 8 out of 10 of the most popular social games today are based on the previous Adobe Flash platform. That world, much like the world of browsers pre-2011, was one in which when you wanted to do something visual, graphical, that felt like an application, you used the CPU. You didn't use graphics to do it. It looked like it. It paled in comparison to what game developers were doing who were accessing the GPU.
Stage 3D in Flash 11 changes all that. We've been working very closely with Adobe on this. It's probably, again, better to show rather than tell. We're going to pull up a Tanks 2 demo that Adobe, it's going to be on the main screen overhead, that Adobe is using as a proof of concept of how we're going to be able to deliver with the next generation of social games or even, you know, games like this, this being more of a combat game, a game console type experience with that new generation of social games and internet games compared to what previously was. For any of you who've played, I won't name games, but some of those current popular social games today, you know that it's a pretty limited 2D kind of an experience.
The key here is Flash is, it's almost a ubiquitous application.
Absolutely.
Every PC, or I think it said on the slide, 98% of the PCs connected to the internet have downloaded Flash. It's really accelerating with the GPU a mainstream application.
Terrific. Okay, thank you guys. Okay, we just took a look at Dirt 3 and saw that was the game that we were playing, DirectX 11 game with Eyefinity 5. Hey, you just got a little bit of data on how our Llano products currently perform versus the competitive offering.
Wait, wait, wait. You're telling me that the APU can play Dirt 3?
Yes, Matt. You're a great straight man. Yes, I am telling you that. Llano could play Dirt 3 without any difficulty, you know, the A-series APU. There's been a, I'm going to paraphrase Bono for a second. Yeah, Bono. I'm surprising you with this one, aren't I?
Yes, you are.
There has been a lot of talk about this next demo.
We didn't promise this, John.
Yeah, no. I'm going to show you guys an F1 game, which is by the same engine, the Codemasters guys who do Dirt 3. This demo, for some reason, got a lot of attention at CES, not from AMD, but from another player in the industry. This is a Trinity notebook that I have up on stage. During the break, you can come and play a little bit more. It's at 13 by 7 settings. That's the settings on this display. We've got the settings dialed up to max, though, so you've got the maximum capabilities. This is our Trinity product that we're launching in. I'm not sure if I'm in the way here of the, are you guys seeing things up there? Yeah.
It's good.
It's time to drive, isn't it?
Yes, it is.
I'm not quite as good at F1 as I am at Dirt 3, but I think you guys can know I would not drive like that and record it for video. That's a good example of what we're doing with the DirectX 11 capabilities that come with every Trinity APU that we'll ship a little bit later this year.
You know, one thing we've proven here today is clearly I'm a better driver.
That's challenging. I'm standing up here. The game's down there. Let me do Dirt 3 again. Okay, it's not DirectX 11 and those direct compute capabilities that came with it with Windows 7. As Matt indicated, with Windows 8, we moved to DirectX 11.1 and some additional capabilities. This is a very innovative company called Unlimited Realities. They're based in New Zealand, a fantastic set of software innovators. They're very interested in the user interface. They're using DirectX 11 capabilities. Think about something like a big all-in-one in your kitchen or your home office or even a tablet device that you might be using.
What you see in the top visual is that using DirectX 11, when you're doing something like playing, in this case, a children's educational game on that type of a device, you can have so much more brilliant effects and lighting and visuals out of the same content when it supports DX11 versus when it doesn't. On the bottom, from a compute and physics standpoint, you think about with multi-touch, all the ways that we play games that are enabled by or enhanced by touch. This is a very cool demonstration that shows when you try to manipulate an object on that screen, it physically reacts to you. It compresses the harder you press on the finger, you pull back, it snaps back into shape and wiggles like Jell-O. Pretty cool.
Very nice.
Yes.
Sony Vegas Pro. Sony, I talked about video editing with my daughter earlier. Sony Vegas Pro is an enthusiast professional video editing application. It does some very cool things from editing 3D stereoscopic video to other special effects. It now has OpenCL acceleration for GPU. What does that mean? It means that when you do a video preview, for example, it's 5x faster. I talked earlier about doing this with my daughter. We have to wait for the thing to render. Bam, 5x faster because we're using the power of the GPU and worked with Sony to make that happen.
That means getting technology out of the way of the video content creation experience. You want to make a change, you want to see the effect of a new filter or lighting, you can instantaneously see that rendered versus having to do it, go away, come back, and take a look at what you've achieved.
Yeah, absolutely.
We're going to do another on-stage demo. We'll be going to the camera up overhead. This is a company that we have been partnering with for, let's see, going on about a year. They're a company we've invested in through our Fusion Fund, which again, Manju will talk a little bit about more. Here, we can cover a little bit of territory at once with this demonstration. I've got a Z-series APU off of that Brazos architecture powered Windows 8 tablet. This is the Windows 8 developer build that Microsoft issued at the Build conference last year. You can see I'm in the familiar Windows 8, now becoming familiar Windows 8 Metro user interface. We're very excited about Windows 8. We think of it as the world's first graphics accelerated or APU accelerated operating system.
What's more, when you think about the universe of applications that will be built on top of the Metro UI and Microsoft's app strategy there, if you think about tools like HTML5, those lend themselves again to graphics to GPU acceleration. We expect to deliver really a stellar experience on all of those different types of new generation of Metro apps that we'll see later this year. I'm talking about BlueStacks right now specifically. I'm sure many of you use Android smartphones. Some of you may use Android tablets. You have some applications that are your favorites. BlueStacks operates as an emulation layer between the Windows operating system and those Android applications. Again, kind of like the demo earlier, it's a best-of-both-worlds concept where through their launcher, you can maintain all the different devices where you want to access Android applications.
You can maintain clear order for what are the most popular applications you want to use. I'm going to start the BlueStacks launcher. I've moved into their launchpad. I'm going to click on one of the more popular Android applications called Pulse, which is a news reader, news aggregator. It comes with photos. Boom, I'm now using a very popular Android application. You saw how responsive it was on top of Windows 8, response to multi-touch, et cetera. That's a great example again of focusing on the experience and where we can get technology out of the way and ecosystem issues and focus on what the end user would really like to see.
Did you guys see that OK? Did I hold that steady enough? Very good.
OK, that must have been demo number eight.
It was.
Because this is demo number nine. This is AMD SteadyVideo. Most of what we've shown so far have been software experiences delivered by independent software partners that we engage with. This next one is an example of what the thousand-plus AMD software engineers did. We have a very strong software engineering community at AMD. What they did when they were inspired by the APU. It is a technology that comes with our drivers for our APUs and our discrete graphics. You'll find it in both our Radeon graphics family and in our APUs. We will continue to iterate on this product and improve its capabilities. Let me show it to you, and then I'll describe it to you a little bit. Again, it is a demo that's made possible through that GPU computing, that massive amount of parallelism.
The other day, I can tell you that a weird weather formation came over my house in Austin. I was all alone, and I wanted to capture it. I thought it was very cool and looked kind of creepy. I held up my smartphone, and I tried to shoot a video of it. I tried to hold my hand as absolutely steady as I could. When I got back and dumped it into my PC, the whole weather formation looked like this the whole time. I defy you to try to shoot a steady video by holding a camera in your hand. Unfortunately, the internet is filled with just these incredible moments that people have captured with these handheld devices. We have to suffer through watching these things shake and stutter and do all these things that are very distracting and can even give you a headache.
What we're showcasing right now is how we, in real time through the driver, as you are watching these videos, automatically sense the shake and stabilize it so you have a much more enjoyable experience. Think about earlier when I talked about the cloud computing trend. This is an example of the kind of processing we can pack into form factors like tablets and convertibles to give you a better overall experience when you're just streaming that content.
The key is, when you're doing something like this, obviously the one on the left is not steady video. It's shaking around.
Yeah, very obvious here.
On the right, it's automatic. It's something you don't even know is there, but you have a better experience.
We've jumped back out now into, I think this is Internet Explorer 9.
That's good. It's not too long, aren't you?
Yeah. We've jumped back out in Internet Explorer 9. We wanted you to see this while it's in the canned video that we're actually working from, watching in this case a YouTube video. Support for this you will find both in Internet Explorer 9, Google Chrome, and Mozilla Firefox, as well as your local content that you're playing back so that you can get rid of that shaky experience, for example, when you just loaded video in that you shot with your smartphone. OK, great. Thanks, guys.
All right.
Bring us home, Matt.
Nine demos. They all worked.
Yeah, I'm happy.
It looked great. We talked about trends, and we heard this morning Rory talked about going to where the puck's going. I always like when Rory does hockey analogies because I'm Canadian. We need to look at those trends, and we need to take our technology and use that to capitalize on those trends. We have technology leadership. We have technologies that we can use to take advantage of those trends and really shoot ahead of the puck. We want to enable immersive experiences. We want to draw you in to whatever you're doing, whether it's a game or whether it's an application. We want you to forget that that video was shaky. We want you to just see it smooth, not even know that anything's going on. We want to immerse you in that.
When we're doing video editing, we don't want you to worry about how long that's going to render. We want it to happen instantaneously. That's the type of experience that we want to deliver. We have to have applications, and applications really put the bow on the whole present. It brings the whole thing together. We have to work closely with our developer partners to make sure we bring these new applications, taking advantage of this technology and capitalizing on the trends.
that was well said. We still have a little less than 10 minutes before we transition to the third and final breakout session, which means we have time for Q&A. I think we've got our microphones and our team set up to bring those through the audience. While we're waiting for the team to get set up, I just want to thank the AMD demo team who are backstage and Eric up here off of, what is that, stage left. Those guys basically had 30 minutes to do all that setup because none of those could be on stage before and make those transitions. They did a fantastic job. I wanted to thank them.
Nice job, guys.
When I look at your roadmap down the future, this is going to be a more general purpose computing platform, this GPU thing. I was wondering what product generation would you be supporting things like error correcting memory, those sort of things that you expect from a CPU engine or servers.
Well.
You need those. Your Green Team across the street has sort of been doing that since the last generation.
There was a reference to Green Team, of course. We have high-end CPU capabilities, low-power CPU capabilities. We have that. Since you said Green Team, do you want to take that one, Matt?
I think, you know, obviously not talking about specific features. We're going to look at the requirements and the markets where we're going into and add those features as required. You mentioned that it becomes more of a GPU thing. It's really, and there's going to be a discussion later with Phil about Heterogeneous System Architecture (HSA). It's really this heterogeneous compute is the architecture that we see going forward. That's really the GPU working closely with the APU. What will happen with that architecture going forward is they'll be sharing memory spaces and that sort of thing to make them work together more closely and really even improve the experiences further.
Hi there.
Hi.
Talking about natural user interfaces, you gave the example using cameras and Kinect, for example, and so forth. As we look to improvements in voice, as we look to increasingly mobile devices with tons of sensors increasingly going into them, lots of potential. Inputs.
to leverage, and then ultimately, of course, in the cloud to leverage as well to serve a better experience. How do you see either AMD's APU or specifically the GPU at either accelerating some of where we're going with user interfaces or enabling altogether? I think where we are today, we're at a really good stage one for a lot of these new UIs, but it's got a long way to go. It'd be great to get some more color on how you see that evolving.
Yeah, I'll take a first stab at that. I think what you just gave is the situation analysis for what we're looking at for the design point of our future APUs. The elements where Lisa Su closed today and she talked about kind of the beyond 2013 horizon, all of those are examples. Remember, she showed all those next generation of user interface type capabilities and security and those types of things. We view many, if not, well, everything that she represented there, we view as acceleratable by our technology. The UI element is a very obvious one. There's a number of things with security as well. That really is governing our approach to designing the next generation, largely also of our low power and ultra low power products as well.
I think you said it well. The bottom line is the GPU is good at accelerating those types of activities. By putting those technologies together and working closer with the CPU, we'll see that we'll get to that time when all those things like face recognition, biometrics, all that stuff will be happening.
I think one other thing to add, you might remember from Mark Papermaster's presentation where he showed the CPU and the GPU blocks, and then he talked about both other custom AMD hardware acceleration. He even made some reference to there could be third-party hardware acceleration in there as well. The CPU and the GPU may not always be the best answer to every one of those new types of workloads. You might be able to dedicate a fraction of a die square area to some kind of a special purpose hardware acceleration to get the job done even better. You see us doing that today as well. We do that with video processing, and we'll do that with some additional types of hardware acceleration with these future APUs. Right here is for next hand.
Hi, can you help us understand a little bit how the ultra-thin form factor is positioned against the ultrabook? I heard the $600- $800 price point mentioned this morning. One of the biggest pushbacks with ultrabooks is the fact that the BOM cost is too high. The specifications are too strict. How should we think about the ultra-thin form factor as opposed to the ultrabook?
That's a good question. I don't want to get into speaking to our competitor, to Intel's specs for their ultrabook platform. I can talk about what our strategy there looks like. You guys might remember, those of you who come to a number of these Financial Analyst Days, I think it was at a Financial Analyst Day in 2009, perhaps, where we first talked about the big push we were going to make into ultra-thin products. We did that. We started creating what we then termed Athlon BGA products. We won a Best of CES award with HP for an ultra-thin notebook. Those were aimed more in that $400 to $500 price point. We'll continue to service those opportunities. We think we'll get more of them with the waterfall from some of those high-end premium ultra-thins to service with Brazos.
Llano, as you guys know, we did not do a BGA implementation of Llano. Participating in the more premium segment there has not been something that we've been in a position to do. I think you got the sense from how much time we spent on it today. We think we have a fantastic value proposition for premium ultra-thins with Trinity. To step into a form factor where you expect to hear great things about battery life and know that we're going to match or perhaps exceed what the other guy is doing with battery life, to be leveraging from Matt's business IP that comes from discrete graphics, special video processing capabilities, new kinds of innovative solutions where you can go from the best of portability to the best of productivity, if you will. We love our positioning.
In terms of price points, yeah, I mean, we think you'll see Trinity silicon that'll be below $600. We think you'll see some designs that'll be certainly above $800. We also know very well where the volumes are in the marketplace today. Where, let's say, the data tells us, the data for many of you in this room, that where consumers expect to buy a Windows PC, a Windows device in retail or in e-tail. That's as we really need to target on hitting those designs in that $600- $800 window. As you know, above that, things really trail off in terms of the opportunity. Other questions?
Oh, is there one back there? Sorry, that's in the blind spot.
Yeah, in the past, discrete graphics was excellent. PC graphics, chipset, integrated graphics was mediocre at best. Phones didn't have graphics. The way you've shown things today, it sounds like the distinction is getting less and less going forward between what you can do in a mainstream product and these discrete graphics. You guys talked about a two-watt product down the road. What kind of graphics should we expect in those kind of products going forward versus the past? Just help us frame the delta, whether it's getting smaller or not. Thanks.
Yeah, that's a great question. I'll take the first step for the edit, but I think you should.
I think it's the first question on the 2 W you may want to comment.
On the two-watt, I mean, obviously, I can't yet characterize something that was on Lisa Su's roadmap in terms of what its graphics capability will be. I think what you could overall expect is that as we maintain this position of leadership in graphics, anything we do, even at ultra low power, that will still be core to how we differentiate. Again, to set up the next session, the idea of Heterogeneous System Architecture means that whatever of that die we devote to the GPU capabilities, we'll have more and more applications that will take advantage of it because of our thought that it's not just about the CPU. We'll get an application uplift benefit from it, and we'll get an overall visual experience uplift from it.
At the high end, I think the idea of the distinction between the APU and the discrete GPU, you heard it loud and clear today, the roadmaps continue unabated for high-end graphics IP. Who knows how many displays Matt's going to come back with?
Let me build on that a little bit. You know, when we talk about that gap, it is narrowing. There are two pieces to our graphics strategy. One is we want to provide the ultimate and the best experience. We're going to continue to do that from a discrete graphics point of view. We've got the five-way Eyefinity . We have Eyefinity 3D. We're going to continue to push that. We're also looking at it a little bit differently. How do we enhance that platform? One of the features we have is dual graphics. We work with the APU and the GPU together, and we get a performance boost when they're put in the same platform together of, say, 1.6x, 1.7x, 1.8x. Really looking at how do we enhance that platform with the discrete graphics.
Yeah, absolutely. In fact, we talked quite a bit about emerging markets today, or high-growth markets. You think about the China market. That has been, I'd say, perhaps our number one market for dual graphics and has really helped us differentiate there. I think we are about out of time. We need to make time for Manju and Phil to come up. Thank you guys very much. Thanks for the great questions.
Pleasure.
Bye.
Good afternoon. In this world of convergence, consumerization, and cloud, the workloads are changing, and they're changing rapidly. We see the workloads being dominated increasingly by media: images, photos, HD video streams, multi-channel audio, 3D graphics, stereo. People are going to want to interact with their media, with their data in new and exciting ways that lead to a fully immersive experience. This is the exciting opportunity ahead of us. This means an explosion in parallel processing and the need to run that parallel processing efficiently. Today, far too much parallel processing gets executed on processes that were not specifically designed for that process. What this means is a waste of power. Wasting power today is unforgivable. We're at a tipping point in parallel processing. The architecture I'm going to describe to you today, the Heterogeneous System Architecture, we also know it as HSA, is a game changer.
It's going to push us past that tipping point and into a new era of power-efficient platforms. As we were designing the Heterogeneous System Architecture, we had several major goals. The primary one was to make the GPU, the parallel processes in the system, easier to program. Make that unprecedented processing capability of the APU as accessible to programmers as the CPU is today. This means high-level languages and not having to write special code for the GPU. As we do this, it dramatically expands the APU's software ecosystem. This is a virtuous cycle that in turn spurs new categories of applications to be developed, consumer applications that are best experienced on this APU and are unleashed by the power efficiency of the new platform.
I'm going to walk you through the start point of the APU today, the future of this heterogeneous platform, the Heterogeneous System Architecture, HSA, what it means, what its features are. We're going to take you year by year through the roadmap, how we deliver the features, what the different bundles of features mean, and the value they bring to the platform as we bring them to market. I'll talk about how we program HSA, how application developers will program it, and why it's so different to the way programs for the GPU and the CPU are written today. This leads to a new command and data flow that's so much more efficient. The software ecosystem that grows around this platform, my colleague Manju Hegde will come up and describe to you when I'm done. The APU, the accelerated processing unit, has already arrived.
It's a great advance on previous platforms. You've already seen in the demand and the sales of our platforms over the last year and a few months with Llano and Brazos that combining the scalar processing of the CPU with the parallel processing of the GPU and bringing it all into a high bandwidth memory adds value to the platform, and it creates demand. How do we make it better going forward? Primarily, we have to make it easier to program. Once we make it easier to program, it also has to be easy for the application developers to optimize and load balance their code, get very rapidly close to the peak performance of the machine. Every year, of course, as we bring out new APUs, we'll make them higher performance, and we'll reduce the power they consume for a given workload.
As we look at bringing out a new architecture, it's informative to look at the history of how we got there. What we see is there have been three major eras of microprocessor development. The first one, the single-core era, lasted decades. It was a golden era where every year or every other year when we got a new silicon node, single-threaded performance just went faster. Things got better without having to change the software. The free ride came to an end. We say we hit the power wall. What this really means is as we added transistors and we added microarchitecture improvements, the power limit in the system meant that we could no longer increase the single-threaded performance. If anything, this curve is rolling over as the TDPs, the power operating points of the platforms, are coming down as consumers and everyone else wants lower power systems.
They want them smaller, thinner, cooler, lighter. Single-threaded performance was no longer the answer. That led us to the multi-core era, where with Moore's law, we were still getting more transistors on each new silicon node. Now we could use them for additional x86 CPUs and run multi-threaded workloads. Within a multi-core SoC, we ran the architecture that had been proven in multi-socket systems, namely the SMP architecture, symmetric multi-processing. This is an era that was relatively short-lived. We rapidly hit the power wall again, where we could no longer increase the performance by adding more cores and trying to scale threaded performance. There were also limitations in how rapidly parallel software came to the multi-core processes, especially in the client space. That leads us to today. Today, we're in the heterogeneous systems era.
We're still getting more transistors on each new silicon node, and we want to bring them to bear in a way that makes a difference in the experience to the end users. The key here is to exploit the abundant data parallelism in these media workloads, and we do it by using the power-efficient parallel processing units that are part of the GPUs. This is a high-leverage play because the GPU has to be there for graphics and does a fantastic job at the graphics and the immersive graphics experience. Those same shaders, the parallel processes in the GPU, are also the ideal processor for parallel computation. Now you'll see in this era, we're still very much at the beginning of the era. There's plenty of room for growth, plenty of room for delivering higher value and more experiences.
We're temporarily constrained by today's programming models and the communications overhead that's involved in starting work on the GPU, transferring data back and forth, and getting results back. The good news is that HSA, the Heterogeneous System Architecture, blows away both of these barriers and lets us ride this curve into the future. It's also instructive to look at how software has been developed in each of these eras. We see a common pattern from era to era, and the pattern is that at the beginning of an era, the initial programming method is most suited to experts and capable of getting the full peak performance of the machine. Very few people have the capability or are willing to spend the time to program in that model.
Over time, the programming abstractions get better, which draws in more programmers, which builds an ecosystem, which draws in more programmers, and then we get another advance in abstraction, and it gets better still. In the single-core CPU era, we went from assembly language programming to structured languages, object-oriented languages, managed languages. At each stage, you give up a little bit of performance for an enormous step function in productivity and a growth in the ecosystem. Similarly, for multi-core programming or SMP programming, we went from programmers programming in threads, which is extremely difficult, to abstractions like OpenMP and task parallel runtimes. In the heterogeneous systems era, we're still early in terms of the software abstractions. People started by programming shaders, and they went to proprietary APIs, then to standard APIs.
As I'm going to show you in further slides, the abstractions will get better until people are programming directly in high-level languages, the same languages they use on the CPU. Now, if we look into the era of heterogeneous computing, we can break it down into three eras again. It started back in 2002 when we first put floating point shaders into the GPU. Some adventurous early programmers realized that they could get at those floating point engines, admittedly through a 3D API, and do things like matrix multiplication. They succeeded in that, and that started off the trend of using the GPU for general-purpose programming. In this era, several companies produced proprietary APIs to enable programmers to get at compute without having to deal with 3D graphics terminology like triangles and textures and geometry.
This was a great start and a great proving ground, a proof of concept that the parallel processing was going to move to the GPU. As long as the interfaces were proprietary, then the software ecosystem can't take off. Next came the standard driver era. This is a much better stage because now there are APIs like OpenCL and Direct Compute where an application developer can create an application and run it across multiple vendors of hardware. They have access to a much bigger range of platforms that their application is going to run on. We have to recognize that even in this era, it really still takes an expert to program the GPU. The languages are subsets of the high-level languages, subsets of C and C++. There are multiple address spaces, separate address spaces for the CPU and the GPU. Data has to move back and forth.
Perhaps the programmer has to decide when that data should move and when it should come back. Also, there are thick stacks of software between the application and actually running the code on the GPU, which leads to inefficiencies unless you have a very large batch of work to offload. That leads us to the architected era. This is heterogeneous system architecture. In this era, we make the GPU into a peer processor to the CPU. This opens up the platform to mainstream programmers. We open up on the GPU the full capabilities of C++ and other high-level languages, no longer a subset. We introduce a unified coherent address space. This means that the CPU and the GPU are using the same addresses. Memory can be allocated on the CPU. A pointer can just be passed across to the GPU, and the GPU can access the same memory.
In this infrastructure, task parallel runtimes, as we were talking about on multi-core processes, can now be extended across the whole platform and run queues on the GPU as well as on the CPUs. We introduce user mode dispatch. This means the application is able to give a task, dispatch a packet directly to the GPU without going through a driver stack or waiting for the operating system. Finally, as we're adding all of this capability to the GPU to act as a peer processor, it needs to be managed and time-sliced just like the CPU is for a fully interactive experience. That is achieved through preemption and context switching. How do we bring all this to market? There are a lot of features here, and we deliver them over several years.
Last year, we already delivered the first stage, which was the physical integration of the CPU and the GPU in silicon with a unified memory controller. You're very familiar with that in the Llano and Brazos products. This year is optimized platforms. A big part of that is bringing out the full C++ language support in the Southern Islands series of GPUs and GPU cores. We also add the ability to do user mode scheduling, where the application can dispatch directly to the GPU. A very big feature is bidirectional power management between the CPU and GPU. This means we can actually slosh power from one side of the chip to the other.
As a workload is heavy on the CPU, we let the CPU cores take all of the TDP of the APU, and as parallel work gets dispatched, that power flows to the GPU as several of the CPU cores shut down. Maybe one is left waiting for results, and now we can use all of the power available on the GPU. This dynamic motion of power within the chip is critical to finishing tasks early, shutting down, and saving power. 2013 is a very exciting year for Heterogeneous System Architecture because this is the year we round out the application features and complete the architectural integration. We unify the address space between the CPU and GPU so they can share pointers or addresses with each other, take results from each other with no copies. The GPU in this era is able to use pageable system memory directly.
We no longer have to pin memory down or have a special area of memory that the GPU operates in. It operates in all of physical memory, and we make the memory fully coherent between the CPU and GPU. This enables new software models like pipelines where data can flow freely between the processes without any need to manually invalidate or flush caches. In 2014, we round out the architecture with the system integration features. These are the features that the operating system needs to do the time slicing and the quality of service guarantees on the system. The GPU compute engine is now able to context switch, and we're able to preempt the graphics and compute. As has been mentioned several times today, we're committed to open standards. We drive open standards and industry standards. We're very active on the standards bodies.
We recognize that open standards are the basis for large ecosystems. Standards win over time over proprietary systems because that's what the software developers desire to have large markets for their products. HSA is an open platform. As we were designing it right from the start, we were always committed to opening it up, publishing the specifications. The specifications are currently out with partners, and we're taking feedback. This is a spec for the virtual ISA for parallel processing, one for the memory model, which controls how the threading works, and one for the system specification of how it all comes together. We're inviting partners to join us in all areas: hardware, operating systems, tools, and applications. We're forming a foundation to guide the architecture and evolve it forward.
This diagram compares how the software works in the driver model, where calls have to go through multiple layers of software to reach the hardware. In the HSA model, we still have a runtime and a kernel mode driver, but they're not used in the high-traffic data paths. Instead, the high-traffic data paths allow the app to go directly to the hardware or through a domain library. There will be many programming models that run on the HSA platform. These are two of the initial ones. KRONOS OpenCL is the premier programming environment for heterogeneous computing today. We're a key contributor to OpenCL, and we're very active at Khronos in guiding its future direction. HSA has features in the architecture that simply make OpenCL more efficient. We eliminate the copies. We allow pointer passing. We minimize the dispatch overhead.
Microsoft C++ AMP is a very exciting development that Microsoft announced at our developer conference in June of last year. It's integrated in Visual Studio and available as part of the Windows 8 Metro build. It addresses the huge population of Visual Studio developers who develop for Windows today. Microsoft has also declared that they're going to make this an open standard beyond Windows. I'm afraid I'm getting over a cold, and this is giving my throat trouble. Let's see if I can at least make it to where I hand over to Manju. A key part of our C++ AMP is it's a very elegant extension of C++. It introduces just two new keywords, restrict and array view. This makes it very simple for developers who already have large bodies of C++ code to just modify particular methods and target them to the GPU for parallel processing.
HSA provides a natural roadmap for relaxing the restrict keyword and opening up all of C++ in future years. This is the future of heterogeneous computing. The architectural path forward is quite clear. We're going to take the programming patterns that have been established for SMP systems and simply migrate them to the heterogeneous world. In SMP systems, multi-core, multi-socket systems, data is shared every day between cores in coherent memory. We're extending that to GPU cores as well. We do it as an open architecture. We publish specifications, and we open source the execution stack in order to jumpstart the industry. This gives us a system with heterogeneous cores working together seamlessly in coherent memory. This knocks down the barriers that were there before that were slowing down parallel processing moving from the CPU to the highly efficient parallel processing cores on the GPU.
We have low latency dispatch, which means that it's very low overhead to kick off a task on the GPU. That makes it possible to run smaller chunks of parallel processing there than has ever been possible before. Finally, no software fault lines. This is really important. Software developers put a lot of investment into their code. They don't want to be told they need to tear that up and do something new to move to a new platform. For OpenCL developers, we're providing an easy path for them to move as new features and extensions get added to OpenCL. Even if they don't make use of the extensions, the existing program just runs faster. For C++ programmers, it couldn't be simpler. They identify the parallel areas of their code. They use the new keywords, recompile, and they have an HSA program ready to go. That's the architecture.
I'd like to invite Manju to come up now and talk to the ecosystem.
We'll do questions right after my part. Phil, come on. OK, Phil, you want to? Yeah, because this is one combined presentation. You can ask him all the questions you want. OK, thank you, Phil. Before I go into my presentation, I want to punctuate Phil's presentation a bit so that you don't miss the importance of what he just said. HSA is a large idea. It's also a very audacious idea. Why so? Consider how software programming standards are typically done. This is true for almost any common industry API, whether it's OpenCL, OpenGL, any of them. This is how the process goes. It's always hardware first. People suppose that the hardware is immutable. You cannot change it. They look at all the platforms that are available. They come up with an API that can address all of them.
In that process, they usually put a few unnatural things into the API. Consider GPU computing today. Whether it's CUDA, OpenCL, it doesn't matter. You have to manage two address spaces. You have to manage two memory subsystems. What does that mean? You can't use a pointer. Every time you change workloads, depending on the nature of the job, you have to incur the penalty of memory copies. That's slow. As a result, despite several years of GPU compute and many, many documented cases of 10x, 50x, 100x improvement, there are only between 100,000 to 200,000 GPU compute developers. If you look at that in perspective, it's a fraction of the 20 million or so software developers worldwide. This is a source of much personal frustration to me. I actually joined AMD a year and a half ago. I came from NVIDIA, where I was GM of CUDA.
For the last several years, I've been trying to promote this GPU computing ecosystem. Frankly, I don't think it can go much further without a radical change. This HSA paradigm actually upends that model. In HSA, what did we say? Software first. We are looking at the software developer and saying, keep doing what you do. Maintain your best practices. If you program in C++, continue doing that. Python, Java, keep doing that. We will change the hardware. We'll put coherent memory between CPU and GPU. Now you don't have to make those copies. It's like a shared virtual memory. We'll make the address space unified. A pointer is a pointer. The GPU can dereference a pointer. One by one, we're just getting rid of all those aggravations, which are the obstacle to software developers embracing heterogeneous compute. You can see why this is a bold move.
As bold as it is, that's not bold enough to call it audacious. That is the second part of what we are saying. Phil mentioned we're going to open the specification. Just think about that. We spent millions of dollars on this, multiple man-years. Yet we're going to open the specification, and we're going to go to partners. We're going to go to the entire ecosystem. We're going to go to competitors and exhort them to adopt this architecture. That is what we're going to do. That is what we just announced. Now look at it from the viewpoint of those 20 million-odd developers. Here it is. We are reducing all those nits, all those aggravations they've had. We're giving them now real access to a very performant, very widely established platform. In addition, we're making it easy for everybody to adopt it. What's not to like?
In fact, that is what we are seeing. We're talking to ISVs. I can tell you there's unanimous approval for this model. That is why we think those partners, those competitors, we will be able to build an ecosystem where they're part of this HSA foundation. That is why it's a bold idea. That is why I think the promise of HSA is in its audacity. With that, I want to come to OpenCL. Why? OpenCL is here today. We are engaged for the last two years in developing the OpenCL ecosystem. Granted, it's not perfect. No GPU compute model today is perfect. It does appeal to a lot of applications that end users want. It's a little painful to do it. That is the pain that we're going to get rid of. It's here today. I know many of you here today are financial analysts.
You have to caveat all your reports with that famous phrase, past performance is no indicator of future success. Luckily, I'm no financial analyst, and I'm not so constrained. One reason why I'm presenting this is I want you to judge me, to judge us by the success we've had in the OpenCL ecosystem. As I mentioned, the HSA ecosystem is going to be much easier than this. It's not going to be easy. Nothing desirable is easy. It's going to be easier. One of the reasons I'm presenting this is for you to see how you can take an ecosystem. When you have an open industry standard, that's a card that's priceless. OpenCL, you can go into the technical details of OpenCL. I think the most important thing about OpenCL is that it's multi-platform. It's truly multi-platform. It's not just lip service.
All those companies in the last line, they have announced in the last year or so. If you look at the last year, there's been a spate of announcements. If I'm a software developer, and I'm a reasonably expert software developer, I know I can write one code base, and it'll at least function on all these platforms. True, today there's some optimization that's platform-specific. I know it will run. We've leveraged that, and that has resonated with the industry. Since you have so many analysts, I want to present to you recent data. This is from a survey done by Evans Data Corporation a few months ago, June of 2011, to be precise. It's a fairly substantial survey, hundreds of developers in each of three geographies: EMEA, North America, and APAC. Look at the results. We highlighted the OpenCL results. The numbers are not small.
The question asked was, if you are doing a new project which requires parallel programming, which parallel API would you use? That's what those percentages are, of a fairly substantial sample size. If you look at it, it's not number one. It's number two, number two, and number three. Look at the API above it, APIs which are much older, OpenMP, thread building blocks. These are not really heterogeneous APIs. These are for symmetric, for multi-core processors. Truly, if you look at a heterogeneous API, OpenCL is number one. It's done this in the last couple of years. OpenCL existed only for about three years, December 2008. We have been the main company promoting it the longest. There's definitely other partners and companies promoting it, NVIDIA. If you track this, this is a couple of years' worth of effort.
If you look at actual applications in the market, these are applications that span the spectrum, consumer applications, some workstation applications. All of these use OpenCL. There are some pretty good brand names out here. There's Sony with their Vegas Pro. I think you saw the demo earlier. There's ArcSoft. There's Cyberlink. There's Magix, the most widely used, or one of the most widely used editors in Europe. There's Dasso, a workstation CAD software application. This is just all I could fit on a page. Today, actually, I honestly don't know how many OpenCL applications there are. There's probably close to 100. That's a good sign because last year I could have told you, I could have counted on two hands. If you track when these applications actually came to market, it's the standard curve. I think until last April, it was like this.
It took off, kind of coincident with the general perception that OpenCL is actually going to be sticking. OpenCL and HSA, in HSA, as I said, we don't want programmers to have to do anything unnatural. They can use whatever language they want. The ones that we're going to come out first with are C++ AMP, Microsoft's contribution to heterogeneous compute, and of course, OpenCL. The good thing about OpenCL is that if you have an application today that you've optimized with OpenCL, say some of the applications I showed you on the previous slide, this is one of the few times we can promise you a free lunch and deliver it. The avoidance of wasteful copies will go. We'll take care of that in the runtime. The low latency dispatch, that'll help all these applications. These applications, without any change from the ISV, will benefit.
The beauty is that OpenCL itself is evolving in KRONOS. It's evolving to a space where it's becoming easier for the software developer to use. For those developers who are used to, and those expert developers using it today, and the others that will come up, they can use HSA really well. The hardware features that we are talking about are universal. By no means should you take HSA as comparative to OpenCL. It's the underlying platform on which OpenCL will shine and shine much brighter than it has to date. What is the vision for HSA? People always ask, what are the killer applications? If you look at many things today, I think somebody asked a question in the previous session, what about user interface, new user interface? Yeah, it's there today. Microsoft with its Kinect, you can keep waving your hands and make things move on the screen.
It's hardly where it could be. The beauty of that is with HSA, it can be much better. We've taken all these biometric recognition, augmented reality. You saw a snippet of what video conferencing can be like. All of these, we've analyzed it through our engineering team has analyzed it. They all will benefit from those features that Phil described as part of HSA. I'll give you a couple of examples. Let's take gesture recognition. I mean, in gesture recognition, you can wave your hands. There are several companies working on just moving your fingers, so a finer grain gesture recognition. You don't want big gestures. You want to just move things with your fingers, especially with, say, tablets. If you look at it, one is to get the map of the hand, and that a depth camera will do. There are other technologies coming up.
The next thing is to do the motion estimation, and motion estimation is a very good workload for HSA. Let me concentrate on the third thing. If you really want to show your fingers off, then what you have to do is what they call the convex file, which is a geometric term. That's not easy to speed up in heterogeneous compute, but it has been done. There are papers showing you can get 10x- 15x, but it's just not easy. Very few people do it. Why is it not easy? Because it's one of those things where the load sometimes is good for CPU serial, sometimes it's good for GPU. That's exactly the problem HSA solves with the coherent memory. You don't have to worry about that. Therefore, that becomes an enabling technology for natural user interface where you have to go to the granularity of fingers.
Another example, biometric recognition, face recognition. Everybody's been dreaming of getting face recognition done well in these low-power clients. In face recognition, the standard algorithm is something called a HAR algorithm. It's a multi-stage algorithm. What it does is it looks at the whole scene, zooms out, zooms out about 10% each time. It's about 20 to 22 stages, but the first few stages are really parallelizable because you have a lot of data, so there's enough to keep the GPU busy. As you get towards the end, you've whittled it down, so you can't keep a GPU busy. Now you've got to use the CPU. In today's face recognition, and we worked with a few companies, having to transfer that workload is the killer, the overhead with that. Again, with HSA, the coherent memory subsystem, you get it for free.
We've analyzed all the use cases in these categories of applications. These are all the applications which really become viable, which really become friendly to the user rather than just have a prototype, which is how I see the space today. These will be what drive software developers and the entire ecosystem to HSA. While that's going to be a couple of years out, we have Windows 8 coming out this year. Why I think Windows 8 is really good for HSA is because it's a strong statement from Microsoft that GPU, graphics, and heterogeneous compute is interesting. It's important to them. If you look at the Metro UI, it's natively accelerated by the hardware graphics. If you look at C++ AMP, that's going to be even available on Windows 8. That's Microsoft's big play in heterogeneous compute. In fact, similar ideas.
They're saying, do heterogeneous compute transparently to the developer. That Lambda function will take care of the hints to the compiler. As Phil pointed out, our first compiler for the HSA architecture is going to be C++ AMP. Then HTML5, again, natively accelerated on GPU. Having an industry player of the stature and the power of Microsoft supporting us this year is going to be a strong message, very well aligned with the direction we want to take the industry in. Last year, we had our first AMD Fusion Developer Summit. It was extremely successful. In fact, you may not see this, but Phil and I are still basking in the afterglow of that. We're really looking forward to this year's Fusion Summit. I wanted to give you a sneak preview of some of the top names that are coming. We've got the Chief Software Architect of Adobe, Tom Malloy.
We've got Phil going to lead off the proceedings again. He'll get into much more depth this year. We have the CTO of Cloudera , Amr Awadallah. There was a nice article on him, I think, in The Wall Street Journal yesterday. Read it. Very interesting character, came from Yahoo last. Our Mark Papermaster, who joined us recently and was on the stage today. There are going to be a few more keynotes. These are the ones that we've got the details of what they're speaking. Over the next few days, maybe the next couple of weeks, we'll be announcing a few more keynote speakers. As you can see, there is definitely a hunger and awareness in the industry that this is the direction that we all have to go. To conclude, I want to reinforce my messages. It's game-changing. It's audacious. It's large.
GPU compute, even with OpenCL and CUDA, is reaching a tipping point. This is the thing that we believe will really tip it over. It'll scale from a couple of hundred thousand to a few million. That's the rising tide that will lift all of us. It's applicable not only to the traditional PC space, to all spaces. I showed you some of the applications, even to the low-power space. As you can see from the kind of membership in the Fusion Developer Summit, we are embarked upon winning the hearts and minds of developers because in 2012 and beyond, if you want to be a hardware, a viable hardware business, you better win the hearts and minds of the software developer community. That's what I wanted to say. Now let me invite Phil on stage.
While Phil is coming up here, let me just tell you one comment before you ask your questions. In my job, I like to characterize it in many ways. One way I like to characterize it, especially with respect to HSA, is it's very simple. My goal is to make Phil a rockstar. I think I've got good material. He's tall, he's slender, he's good-looking, and he's British.
Yes, maybe the clarinet.
You want to go back?
What?
You said you want a slide.
Do you remember you had three columns of boxes? Each column had three boxes: 11, 12, 13. It's where you introduced the coherent memory model and the pointers and so forth?
Yes, I remember that.
That suggests, but you never said the words, that there was going to be a load balancer somewhere. Because otherwise, that's damn good work. How do you set up these pagable systems on this coherent pointer and not do load balancing? Or is the programmer still going to have to explicitly assign what processor gets what string?
OK, so the expectation is still that the application developer knows their application better than a runtime or a driver. We're not looking to automatically load balance. We're not looking to create a system that can examine the characteristics of the program while it's running and move it from, say, a CPU to a GPU or from a GPU to a CPU. We're still looking for the developer to say, OK, this method has parallel work. We'd like to make it a candidate for the GPU, and we want it to execute there. We actually think it would be a mistake for us to try and build a thick runtime that does that automatically. We're still relying on the developers to do the right thing. We just want to make it extremely easy for them, so they effectively just mark the method, and then we take care of the rest.
We're making the programmer guide the compiler.
Yeah, yeah. I mean, because today, if you look at it, we really make it difficult. Not only is it difficult, but it's very inefficient. We're getting rid of those two aspects. Ultimately, the programmer knows his program better than the driver or the runtime.
I want to follow up on the HSA Foundation that you mentioned. In particular, what's your intent there? Are you going to hand off? Is it going to be like a standards body that's going to be governing all this? AMD then will have to do whatever that foundation comes up with? Is it going to be a board of directors that's going to guide AMD? How do you envision that working?
Yeah, if I wanted you to know that level of detail, I would have put it in my presentation. My point is we're not trying to invent anything new there. I mean, there are tons of foundations. In fact, we picked the word foundation because if you look at Linux Foundation, Apache Software Foundation, these are thriving, very important software bodies that drive open standards. We are going to use one of these models. We are actually looking at a ton of them to make sure that it's very low overhead because we want to have many companies join it. It is going to be nothing unnatural. It's just that today we don't want to go into the details of that.
Yeah, the important thing is we want to make sure that the partners who join us in this have a voice in the architecture going forward. It would be foolish of people to join in on this architecture if we maintain full control of it. We are getting the ball rolling, getting the initial specification together, and taking a ton of input from partners. Once the architecture is complete, we want a body to take it forward where everybody has a voice.
How do you define when it's complete so you can hand it off?
For HSA, we're looking at the feature set you see on the screen as the 1.0 complete feature set. There will be compliance tests written to this, and this will be the 1.0 standard. That's the natural place to hand it off. If you try and do a standard from a standards body from scratch, it can take you five years to get to 1.0, and it can be something unrecognizable at the end of it. By taking it with stewardship through to the 1.0 level, we think we get the best of both worlds. We get to market quickly, we still take a lot of input, and then set up a situation where everybody has a voice in how it goes forward.
Unless you think that this is a very peremptory set of features, we have taken input from a number of companies, all of which you probably know very well. Their inputs have been incorporated into the current version.
To that end, I'll follow up. How are you working with some of your competitors? Does this HSA model extend beyond x86 to other, for example, ARM cores along with GPUs? First of all, specifically, obviously, the elephant in the room is Intel. Is Intel necessarily going to be that interested in this if this actually gives you, in the end, a competitive advantage? If they aren't, then doesn't that put a halt to it early on?
Yeah, if you want something to be an open standard, it would be foolish to ignore some of the ISAs that are popular. The best argument, whether it's Intel or any company X, is the argument I made. In 2012 and beyond, you want to succeed in any place, you better get the developers excited. The story, as I said, is very powerful. All the crud that has prevented developers from accessing heterogeneous compute, we're resolving that. By the way, this is not just our opinion. As we talk to partners, the most common response is you solved our problems for us. We are making it open. I don't want to address Intel specifically. For any company that wants to really lead in this space, they would have to solve the developer problems.
There is solving the developer problems, and then there are business practicalities. I guess that's kind of what I'm trying to get at, if you can comment on that.
Yeah, when you take something into a standards body and make it open, what you're doing is you're saying that you will allow some innovation. The more important thing is standardization. That's exactly the model we will follow. It's not going to be a standard like an RFC in the internet space. There's going to be room for innovation. Fundamentally, we're going to make it easy for developers to access the power of the platform. Why would we do it? Our answer is simple. A rising tide lifts all. We've got, I don't remember, 80 million- 100 million GPUs last year. Most of them, except when those guys are playing games, are lying fallow because there are only 100,000 people who can really do something with them. Let's take it to 5 million.
That's a much better situation for us because we know that in the heterogeneous IP, the leadership that Mark talked about, that's a contest that we are very comfortable winning.
If I could just add to that, I tried to make clear in the slides. The way the HSA specifications are set up, it is ISA-independent on both CPU and GPU. Yes, it will work on another ISA. If I think through the list of companies we've dealt with already to take input on the specification, it includes many competitors. We're not in a position to disclose names of specific companies today. Competitors have provided input into this spec. We've modified the spec, and it contains their input now.
Phil, what is the memory ordering model for HSA? Would you describe it as strong or weak? How do you plan to resolve the difference between instruction sets that are weakly ordered and strongly ordered? Because even if you can use any ISA, some may not produce the same results or correct results.
OK, the way the memory model for HSA works is it's weakly ordered, but with strong primitives for synchronization. It's basically a load acquire, store release, and barrier model, if you're familiar with those details. Now, because it's weakly ordered with synchronization, any ISA that's either weakly or strongly ordered can be compliant. A strongly ordered memory model can always meet the constraints of a weakly ordered memory model. It's not the other way around. We have been very careful right from the beginning to completely specify the memory model so that as platforms come out from different companies or even the same company year to year, they are consistent and compliant with that memory model. We will have tests to that memory model. We're confident that all platforms can address this. It hasn't been done in a way that favors one implementation or another.
Doesn't that mean if it's weakly ordered, you can have different results? Like you can get a different answer if you run it through a strongly ordered system like an x86 CPU versus a weakly ordered system like, say, ARM?
No, I don't believe so. A correctly formed program will give the right results either way. If the program is not properly formed, in other words, the thread is trying to access results before a synchronization object has been signaled, it's possible to get different results on different implementations, whether they're weakly ordered or strongly ordered. A correctly formed program would not give a wrong result on a different platform.
All right, thanks. Intel has used a concept of using not massively parallel, but parallel x86 small, x86 core as a co-processor in the server market. Is that an alternative approach to HSA or heterogeneous computing?
I don't see it as an alternative approach to HSA . I think it's solving a different problem. You can almost look at it as a different approach to parallel computing in a legacy model, right? Where it's a different approach, for instance, than using a GPU. There is nothing that would prevent such an architecture adding features and becoming compliant with HSA. In fact, that would be an interesting direction.
I think the word you were searching for was MIC. Many integrated.
Thank you. Two questions. First of all, just trying to understand the AMD Fusion architecture here and how that compares with someone using a dedicated video codec or DSP or touch controller. Are you targeting a certain class of problems versus those SoCs that are out there? There's a performance trade-off as you do that versus using these probably more power-efficient codecs and so forth. Second question is, can you envision?
Can I take the first one before I forget it? It's a complicated question.
Go for it.
OK, so what we're defining here is a platform and a memory system and how the different processes in the platform execute and how they share data and interoperate with each other. Within these slides, I dealt with CPU cores and GPU cores. However, the platform naturally extends to other forms of processors like DSPs, codecs, fixed function units. They wouldn't necessarily use HCL because they're not compiling through that parallel ISA. In terms of sharing memory with common pointers and coherency and context switching, all of those apply. The platform naturally extends to embrace those kinds of processes. The other comment I would have is, as we go forward with SoCs in a very power-demanding environment, we're going to see more and more black silicon.
In other words, SoCs with a variety of processes, CPU cores, parallel cores in GPUs, fixed function processes for particular codecs, and other use cases, all of which interoperate in memory. If they're not in use, they're power-gated off. That's what makes it black silicon or dark silicon. It's a very natural way to go forward. When you want to build a piece of silicon for one specific workload, like a codec, the matched processor will always be the most power-efficient method. It becomes a trade-off for us or an OEM of whether they want to spend silicon for one particular workload or whether they'd rather be able to run multiple different workloads across shared processors like the parallel processor in the GPU. Going forward, I expect it to be a mixture. For instance, for the most popular codecs, you may have fixed function.
For newly emerging codecs or legacy codecs that are getting less use, they'll run on the very power-efficient GPU cores. Next part of your question.
Sure. And, thank you for that. The second part was, do you envision systems that could be using an x86 and an ARM processor? I mean, there are systems that do that, but in terms of more integrated going forward. Thank you.
Your question, I assume, is whether I imagine a system that has both x86 and ARM cores in it working together. That's the question? The architecture supports that. I don't necessarily see a lot of business cases for sort of having matched capability processes at different ISAs ping-ponging together, but conceptually, it would work.
I guess my question was, would there be a value-add to doing that versus the x86 or ARM separately? Thank you.
I'm not immediately seeing the value of that, but as I said, the platform is architected to support it.
Now, making any assumptions about the bittedness of the past CPUs that would be supported within HSA, i.e., 32 versus 64-bit?
Yeah, actually, we've thought a lot about that and made some decisions on it. The platform will naturally excel with 64-bit processing. We're seeing burgeoning amounts of data. Even in mobile devices, smartphones, tablets, I see people hitting the 32-bit barrier very, very soon. It would be wonderful to make this 64-bit only and simplify the architecture. The reality is, as we talk to partners, there is a strong demand for 32-bit versions of HSA, and we will support it.
Now, would you support both mixed modes? Or is there going to be a 32-bit version and a 64-bit version? I mean, if I'm an ISV and I want to distribute something that I've architected for HSA, am I going to have to release one version for 32-bit machines and a separate version for 64?
You're not going to have to do that. I mean, you could, for instance, release just a 32-bit version of your application that would run both on the 32-bit and the 64-bit platform. The likelihood is that application developers are going to want to take advantage of that 64-bit address space. In that case, they may have to do two versions.
Great. Thank you.