Graphics Processing Units: Prosperity or Extinction?
Graphics processing units have been gaining importance ever since the first 3D games that could use their functionality emerged. But will discrete GPUs actually exist a decade from now, or will they share the fate with standalone audio cards or add-on floating point processing units (FPUs)?
With the release of hybrid processing units with x86 and graphics engines inside the market of low-end and inexpensive mainstream graphics cards in general will inevitably shrink rather considerably, especially in the mobile segment, since both Intel Sandy Bridge and AMD Llano promise very high performance in video games and other GPU-accelerated applications. In order to keep the segment of graphics boards that cost less than $100 alive companies like AMD and Nvidia will have to offer very advanced products that offer major performance difference compared to integrated graphics products. Naturally, this is not a matter of technology, but rather a matter of economics.
Recently Nvidia outlined its plans till 2013 - 2014 and we do know that ATI, graphics business unit of AMD, has plans for at least two more generations after the Radeon HD 6000. Nvidia even outlined its design goals for the forthcoming Kepler and Maxwell architectures, it is not doubling gaming performance every year, but it is aggressive improvement of double precision floating point performance per watt, a clear indication that the company sees general-purpose and high-performance computing on GPUs something tremendously important.
There is a definite consensus between AMD and Nvidia that graphics chips are very suitable for highly-parallelized computing. Both companies have HPC GPU deployments and it is clear that the market of GPU-accelerated high-performance computing will only grow. That said, both companies will continue to develop new graphics processors with HPC and eventually cloud computing in mind.
But the market of HPC should not be overestimated. According to Mercury Research, the total available market of discrete graphics processors (desktop and mobile) was approximately 34.9 million units. The number translates to TAM of approximately 100+ million of GPUs per year. The world's most powerful supercomputer today - the Jaguar - utilizes around 37 000 of AMD Opteron microprocessors with six cores. The K supercomputer jointly developed by Fujitsu and Japanese ministry of education will feature 80 000 thousand of SPARC64 VIIIfx chips with eight cores to deliver 10 PetaFLOPS of double-precision performance. It takes many months or even several years to assemble giant supercomputers. Since supercomputers are usually based on dual-socket blades and it is possible to install maximum two compute boards into those blades, it is not hard to count that the TAM for HPC compute boards will remain in hundreds of thousands, maybe millions, in the coming years, but not tens of millions.
In fact, the world's two largest HPC GPU deployments are Nebulae supercomputer with 4640 Tesla C2050 cards as well as Tianhe-1 that sports a total of 2560 dual-chip ATI Radeon HD 4870 X2 cards (5120 GPUs). Still, the market is growing fast and at some point, maybe in 2020, the TAM for GPUs in the HPC space will be comparable to the TAM of graphics chips today. Of course, if Intel and AMD manage to release server processors with integrated MIC or GPU-based accelerators, the interest towards specialized accelerators will start to decline.
A server with two AMD FireStream inside
For PC gamers continuous development of GPUs for the HPC market means that GPUs will not become a niche product, but will continue to exist and evolve even in 2020. Another reason why graphics processors should continue gaining graphics performance is inevitable increase of resolutions and inability of integrated graphics products to quickly gain performance to respond to demands of the consumers on the leading edge of progress.
However, the market of graphics solutions in a decade from now will be completely different from 2010. Inexpensive graphics chips will have to offer much higher levels of performance related to integrated graphics than today and maybe even have certain exclusive features. Graphics cards of the day after tomorrow will not only be in add-on form-factor as further popularization of notebooks will drive demand towards external graphics cards. In the future it will be possible to plug an external graphics card to a laptop with decent microprocessor and play latest video games wirelessly on a large screen. Perhaps, graphics rendering will move to server (which still means that the server will need to have a GPU) and hence premium graphics experience will be possible even on low-power handheld devices.
What can we expect from GPUs ten years down the road? In case the evolution of graphics chips will be as rapid from 2010 till 2020 as it was from 2000 to 2010, then this is a question that no one can answers. The main thing that we would expect is consolidation of graphics and compute APIs in order to simplify programming of different applications. However, Jon Peddie, from a leading firm that analyzes graphics market, does not believe this and claims that graphics-specific APIs will continue to exist in 2020.
"[Graphics APIs] will be necessary to support the unique architecture of a GPU," said Mr. Peddie.