AMD predicts multicore future

You can never underestimate the competition...We are conscious that we have some core technologies we want to get to market quickly, and we can't take our eye off that.

As Advanced Micro Devices Inc. sees it, the next big leap in mainstream computing is all about marrying different kinds of cores on a single die. Creating some unique products in that space was one of the main motivations behind the $5.4 billion merger last year between AMD and ATI Technologies. Already in the pipeline are the so-called Fusion processors, which will include a variety of CPU and graphics cores. The first Fusion chips, initially targeted at notebook PCs, are expected to hit the market in 2009.

Dave Orton, the former CEO of ATI who now heads AMD's graphic division, sketched out some of the targets and technical hurdles that the combined company has faced in getting Fusion devices ready for market. Orton spoke with EE Times' Rick Merritt and also offered an update on AMD's vision and its progress toward this multicore future.

EE Times: What will the end-user see in a Fusion notebook?
Dave Orton: It's just what ASICs provided when they emerged in the 1970s—they provided cost, power, quality and area advantages. Less area means fewer pins, less I/O and lower power. Higher integration means less cost, and that means the motherboard can be smaller. So you have a more portable, lighter notebook with more battery life. That's the opportunity for Fusion in that part of the market.

As we think about a lower-end Fusion device that could go a little further downmarket, it could be a phenomenal mobile device.

For a smart phone?
Yes, exactly, for those kinds of devices. We see a broad range of possibilities. The challenge now is choosing which ones to pick and which ones can create new categories in the market. That's what we are working on now.

What about Fusion for DTV?
The processing power of a TV will continue to grow. If you look at the software stacks for DTV, there are Open Cable STBs and Media Center systems that are more like PCs. We have the opportunity to bring together the concept of what a PC is and the image processing of what STBs want to have.

So the concept is this: a device that is not a PC slides into the living room, but the insides of the device are like those of a PC because of the software stack, the connection to the Internet and maybe it even has an IPTV play.

Today the Open Cable STB and the media PC represent two different worlds.
They are two different worlds, and there are ways to bring them together.

What graphics cores will you use for Fusion processors?
Depending on the market, the graphics performance and the area that you want to dedicate to graphics will vary. When you think about the notebook market in 2007, it will have to be some level of a Microsoft DirectX 10 core. If you go below that into a high-growth market, it could be a DX9 core. So it will vary.

Are you designing these graphics cores from the ground up or using existing cores?
Probably a blended model. You can always waterfall, but what we have found even in the GPU space is that it's challenging to waterfall the highest-end enthusiast part all the way down to the value segment.

We recognize that we have to either attack two design centers—leveraging as much as we can—or start in the middle and waterfall both ways. For the ultralow-end, we found we have to think from the ground up to get full area optimization. The partitioning of what you run on a CPU and GPU would be slightly different in some of those cores.

So you are designing graphics cores specifically for Fusion?
Cores for Fusion will leverage as much as we can from our GPUs, but we will have to do some optimization. It won't just be a waterfall. You have different libraries, process technologies and partitioning.

Because the CPU and GPU are closer, what you put in them will be slightly different. You have an opportunity to optimize the interfaces and power management.

How do your ideas on multicore play out in your chipsets?
Today, we are bringing separate graphics and video cores into a common silicon environment. Our latest 690 chipset uses what we call a universal video decode core.

There are cores that we leverage from our DTV development as well. We want to make the PC become a more HDTV-capable platform—not just for HD video, but actually for TV. We are also asking what other processing we want to pull into a chipset, especially when you think about I/O processing.

What sorts of I/O cores are you considering?
The interesting ones are wireless, wireless-USB and broadband connections, which are technologies that the company doesn't have inside. That's a good opportunity to partner.

How far do you think will the trend of using graphics for other kinds of applications go?
We will be the first to say that there is a clear set of problems that needs to run on standard CPUs and the x86. The world has tried to do coprocessors since the 1980s.

What we are finding is that there is also a class of data that lines up with threads or streams that map well to GPUs and the general-purpose processing elements now in the graphics chip. They are the classic high-performance computing codes and some commercial codes, like facial recognition or search.

The opportunity is that more kinds of applications will find value in dispatching parts of their code to a GPU. The CPU is still the heart of the system, but you can do things faster, better and more effectively using this combined processing power of CPUs and GPUs.

What are the technical issues bringing the ATI and AMD cores together?
In the fabless world, ATI has optimized largely around area. There are things we have learned that the custom designers can learn from us.

On the other hand, the unrelenting pursuit of performance and number of gates per flops is a concept we know in the GPU world, but we don't live and breathe it the way that CPU designers do. Some of that and some of their power-management techniques we can start to adopt. We also live in very different worlds in terms of time-to-market. CPUs probably have nine months of construction and a year from tape-out to production. GPUs live in a world of three months of construction and six months to production. So there are opportunities for the CPU world to learn from the fabless world.

What are your plans for using the AMD fabs?
Over the next two years, the AMD processor will consume 100 percent of the AMD fabs. As we look at 2009 and 2010 and our fabs become fully configured, we look for a flexible model for fabrication. That will be around the 45nm time frame, when we will have more capacity and bulk CMOS used by ATI, and silicon-on-insulator (SOI) used by AMD will become more similar in their design methodologies.

That's the point where we look at what processors can become fabless. We will make sure there are a few parts that, if the market turns down, we can roll them into the AMD fab in a bulk or bulk-like SOI process to make sure we have well-utilized fabs.

What do you think of reports that Intel is ramping up a new graphics design team and planning processors that also use different kinds of processor cores?
You can never underestimate the competition. As one of the leaders of Intel once said, "Only the paranoid survive." But we are conscious that we have some core technologies we want to get to market quickly, and we can't take our eye off that. We merged to deliver a vision of what we can do in 2009 and 2010.

The engineering teams on both the AMD and ATI sides are excited about what we can do. We had a technology forum with all our key technologists in San Francisco for a couple of days.

What came out of that meeting?
It was the first opportunity to bring together all the fellows and technologists from AMD and ATI. It was a "mind meld" to get everyone exposed to our combined technologies and the opportunity we have to work together in core areas. The output of that work will be centers of expertise—core teams to tackle some of the biggest problems. Over the next nine to 12 months, new innovations will come from these core development teams, thanks to the combination of the best-of-breed ideas from both companies.

What are these centers?
As part of the R&D integration—and I've been heavily involved in the R&D integration—we've kicked off probably 15 centers of excellence in areas such as power, packaging, SoC and Fusion processors.

- Rick Merritt
EE Times