Compute acceleration of highly-parallelized algorithms is indisputably a potentially huge new market for heavily-parallelized graphics processing units (GPUs) AMD and Nvidia Corp. Intel Corp. understands it and is working on its own accelerators for high-performance computing (HPC) space. But Nvidia doubts that without financial backing by the graphics cards market it is economically feasible to develop an HPC-only accelerator. Moreover, given Intel's track-record with Larrabee, Nvidia does not consider Intel's MIC [many Intel core] architecture a threat for its HPC business.
At present Intel is testing its code-named Knights Ferry MIC product (which is actually Larrabee graphics processor) with developers, but the first commercial MIC-based design will be released sometimes in 2012 and is currently known as Knights Corner.
"It [Knights Corner] is so far out right now in terms of when it gets to production and when its gets to the market that I do not consider it to be a threat at this point," said Sumit Gupta, product manager at Nvidia's Tesla business unit, in an interview with X-bit labs.
Both ATI, graphics business unit of Advanced Micro Devices and Nvidia nowadays sell special accelerators for HPC applications that are powered by graphics processing units originally developed to accelerate video games. Since GPUs contain massive amount of stream processors, they turn out to be extremely efficient on highly-parallelized code and offer tremendous advantages compared to traditional multi-core central processing units (CPUs) in terms of performance per watt. After Intel Corp. failed to develop its own discrete graphics processor code-named Larrabee, it said that it would concentrate on creation of many-core chips aimed specifically at HPC market. But according to Nvidia, whose Tesla compute boards power three out of top five supercomputers in the world, the high-performance computing (HPC) is not the primary market for graphics chips. As a result, it hardly makes sense to develop a chip specifically to power supercomputers.
"I do not see the economic model [with MIC]. We are able to produce those [Tesla] GPGPUs because the ultimately there is one GPU for GeForce consumer graphics, Quadro professional business. It costs $500 million to $1 billion to develop those new products every year, it is a hyge investment. Unless you have that [consumer and professional] economic engine in the background, I cannot imagine how one could make a GPU without having a graphics business," said Mr. Gupta.
In the most recent quarter Nvidia earned $210.1 million selling professional Quadro graphics cards and Tesla compute accelerators, $581.9 million selling discrete and integrated consumer GeForce graphics adapters as well as $51.9 million on various consumer electronics products. It is evident that the bulk of the GPU business is in the consumer market. Even though Intel unquestionably can afford development of an HPC-centric chip, it is unlikely that it will pay back in the short-term.
"HPC is a [relatively] small market. We saw back in the past that HPC-only companies came and then vanished. We will see... We have a huge head-start, we have a product shipping, we have all these customers and deployments. Intel will try to catch up with us when it comes out with its product," added product manager at Nvidia's Tesla business unit.
The HPC market is indeed not big for Nvidia. Even the top performing supercomputers - Tianhe-1A, Nebulae and Tsubame 2.0 - utilize 7168, 4640 and 4200 Tesla compute boards, respectively; other systems utilize even fewer GPGPU boards. Meanwhile, Nvidia sells millions of graphics chips for consumers every year.