GPUs, low-power pave path to exascale supercomputing
11/3/2011 12:58 AM EDT
Power, floor space constraints
Despite the money being invested in HPC, however, many potential supercomputer clients are still being constrained by power restrictions or even floor space limitations, said Claunch, adding that he believed tremendous benefits could be gained through increased efficiency.
“We are really taking the power efficiency challenge very seriously,” said Supermicro's Clegg. “Power supplies are being looked at more carefully,” he added, noting that Supermicro aims to get the majority of its power to a platinum level of efficiency, or some 94 percent plus. “Power and cooling is the biggest problem,” Clegg re-iterated, noting that it was becoming increasingly difficult to achieve a favorable cost-to-benefit ratio with cooling costs increasing almost exponentially as performance increased.
“There’s not enough cheap power to get us past the exascale level unless we make some serious architectural changes,” he said.
“Power is the main challenge on the path to exascale computing,” agreed Anthony Kenisky of Appro.
Chuck Moore, AMD corporate Fellow and technology group CTO, said those looking to achieve exascale would have to factor in a million dollars per megawatt. “As good as Bulldozer, or Interlagos is, they are not good enough; they’re not going to get us there," he added.
Moore predicted it may take AMD until at least 2019 or 2020 before its chips would offer a level of programmability sufficient to take customers to exascale level, noting that GPUs would factor heavily into the equation.
Indeed, the majority of panelists agreed that the use of GPUs in supercomputers is becoming an integral part of the segment’s forward momentum, though the consensus was that CPUs would in no way become redundant in the space as a result.
“GPUs are a very important part of heterogeneous computing in terms of alleviating the bottlenecks,” said Clegg adding that graphics processing was “right on the cusp” of becoming accepted as the predominant way to get heterogeneous computing going. While that was a significant achievement for the GPU space, however, Clegg was quick to temper his comments with caution. “Will the space become 100 percent heterogeneous and GPU based? I don’t think so, because there are some applications that are suited to it and some that are not,” he said.
“GPU is in vogue in HPC at the moment,” added Charles Wuischpard, president and CEO of Penguin Computing, which runs a supercomputing on-demand style model. “For our larger systems, everything we’re doing involves a GPU but for the mass market, not so much,” he said.