APU - solution for the future?
The other, more significant example is "Fusion," a code-name for a processor that "fuses" a graphics core with a traditional CPU and the Northbridge on the silicon level. The reasoning behind that idea is that basic 3D graphics is well on its way to becoming a commodity and that graphics processors have many more talents that aren't exploited today: For example, the floating point performance of ATI's R580 processors is already almost ten times higher than that of today's high-end dual-core processors. Technically, the integration of a GPU into a CPU, could enable AMD to build very low-end processors that do not need an integrated graphics chipset anymore - and use the same idea to build supercomputer microprocessors.
AMD recently told us that Fusion is very much in the first stages of development, so its capabilities are largely theoretical at this time. AMD, however reiterated at the meeting with analysts that the first Fusion processors will be mobile processors and will provide advantages in terms of power efficiency when released. Performance is pure speculation at this time, but there is a good chance that the integration into the CPU could eliminate bottlenecks and bring some benefit over the external integrated graphics we are used today. AMD also expects that such processors will be cheaper to produce than separate CPUs and integrated graphics chipsets.
What is enticing about this idea, is the fact that Fusion is basically the same approach that brought us the integrated memory controller in Opteron and Athlon 64. AMD has been quite successful with this approach and integrated graphics are certainly a novel idea that could work out similarly. AMD's roadmap suggests that the first Fusion will surface in the 2009 timeframe.
But AMD does not call Fusion a "GPU-CPU," it calls it "APU" (accelerated processing unit) - which means than just a different way of integrating graphics. In fact, you could call this a "Lego" approach, which enables the company to take different building blocks to create different processors with very specific target applications. AMD claims that this "modular" approach to building processors will enable the company to quickly react to changing market trends. On the higher-end, a GPU could morph into a "stream processor" that won't accelerate graphics but take advantage of the floating point horsepower of the processor. AMD envisions stream computing to move into the mass market at some point, but clearly, stream computing is a high-end topic today and non-existent in the mass-market, as there are no off-the-shelf applications that are taking advantage of floating point capability.
What concerns us about Fusion is how the consumer will be affected, if this idea is successful and will develop as promised by AMD today. The company dreams about building a "one-size-fits-all processor" - and that is true only from a development or manufacturing point of view. A processor that is performance-tailored to certain application scenarios will only reveal that performance in those areas. There is a good chance that we are going from a general purpose microprocessor today to a specialized processor, which means you could end up with multiple computers in your house, specialized on different tasks.
Asked about a possible "fragmentation" of the microprocessor market, AMD told us that "people should not be worried about what a device is supposed to do" and believes that Fusion will be able to make buying a computer - or any other device that could integrate an AMD processor. From today's view, however, we believe that specialized processors could make buying a computer more complicated. At least if the consumer will continue to be exposed to the task of deciding on a processor for his computer.
Is multi-core dead?
Chief technology officer Phil Hester told analysts that getting overboard with multi-core processors could be compared to the Gigahertz race. Of course, this was a shot against Intel, who has been distributing the idea of dozens or hundreds of cores in one processor. When looking back, we realistically have to say, however, that quickly increasing clock speeds were the right strategy around the turn of the century, but the industry should have turned the corner towards more power efficient technologies earlier. No doubt about it.
The multi-core talk today in fact sounds very similar to the Gigahertz trend around 2000. It appears to be solving many of current problems such as high power consumption and promises to increase application performance at the same time. While Intel has been promoting the idea of "many-core" processors, the company appears to have retreated from that strategy at least partially. Current roadmaps from Intel suggest that notebook and desktop PCs will be stuck at quad-core for a couple years, while greater core numbers (8) could be realized on server and workstation processors before the end of the decade. AMD explicitly said it won't participate in such a many-core race.
That claim however, exclusively refers to "homogeneous" multi-core processors with a number of identical cores. "Heterogeneous" cores, with processing cores that could be assigned different tasks are not quite new and have been discussed by Intel in the past - AMD's APU's are heterogeneous multi-core processors that are likely to see an increasing number of cores as well. However, AMD is first to lay out a clear direction of such heterogeneous multi-cores. So, multi-core isn't dead.
Intel, by the way, is also aware of the need of number crunching horsepower: SVP David Perlmutter told us in an interview earlier this year that "integer and floating point performance will improve significantly" in the upcoming 45 nm core due in Q4 2007. Also, 2006 was clearly the year where the technology came out of its shadows: Take for example Clearspeed's CSX600 accelerator card or the PeakStream software - two concepts that show that there is a trend that makes a lot of sense and has great potential.
The software perspective
The real challenge in stream computing, however, may not be building "stream hardware." There aren't many applications that are taking advantage of floating point horsepower today - it is mainly an area that is limited to scientific and financial applications. Mainstream success of such processors will largely depend on the availability of software - and it is unclear how long it will take until developers pick up the trend.
AMD thinks the adoption of stream computing may take about as long as it took developers to embrace 64-bit - two years. But realistically, 64-bit is still a high-end topic and virtually non-existent in the mainstream (at least until we have a need for computers with more than 4 GB of system memory). Intel came out with its 64-bit extensions 2 years after AMD, but still well in time to take part of a 64-bit trend, when it begins to grow.
Turning the microprocessor space into a different direction is a monumental task -certainly much more elaborate than anything AMD has done before. I agree with AMD that stream computing could take the same route as 64-bit, but this one is a different caliber. Having the vision is one side of the story, but making it a reality will be more difficult.
