Executives from some of the largest players in the supercomputing industry gathered in San Francisco yesterday to discuss the future of high performance computing (HPC) at the HPC Industry Roundtable, hosted by AMD and moderated by Carl Claunch, VP & Distinguished Analyst at Gartner.
Panelists included Don Clegg, Vice President of Marketing and Business Development at Supermicro, Charles Wuischpard, President & CEO at Penguin Computing, Margaret Williams, Sr. Vice President of HPC Systems at Cray, and Anthony Kinisky of Appro. Chuck Moore, Corporate Fellow and Technology Group CTO at AMD, also spoke during the discussion, which revolved primarily around the challenges and opportunities facing the industry as it enters the next phase of development in HPC.
Specifically, the panel addressed questions such as: the relative significance of GPU computing, the feasibility of HPC in the cloud in the near future, the role of programmability in driving different architectural iterations, and the role of supercomputing itself in progress on a global scale.
A common concern spanning almost every topic was the energy cost associated with operating such massive computing systems. According to Chuck Moore of AMD, on one level, exascale computing is available today - hypothetically represented in the sum capacity of all computers and networks available - but the real issue is cost. Though upfront costs of commissioning a customized HPC system have largely been eliminated, the increased demand for performance results in astronomical rates of energy consumption. When people talk about exascale computing, what they're really talking about is exascale at 20 megawatts, said Moore, going on to briefly explain that a comparable system operating today would use approximately 200 megawatts yearly at the prohibitive annual cost of $200 million.
Addressing the question of imminent availability of HPC in the cloud, the panel again expressed skepticism about the feasibility of paying for such a system. The biggest challenge of all is, 'How are you going to pay to run this?' said Don Clegg of Supermicro. It's not just about scaling more and more - there's not enough cheap power to get us beyond the exascale unless we do some [work on the] architecture systems.
According to Margaret Wiliams of Cray, an immediate challenge on the software side is improving the resiliency of networks and models in light of the varied sources that contribute the myriad of parts of today's HPC systems. In addition, another key software challenge was ensuring programmers' continued ability to build parallelism functionality into the systems. How do I, as a programmer, express parallelism so it can be mapped? asked Williams, posing the question. How does the application guy figure out how to get through that space to access the compute power behind it? Both Williams and Anthony Kenisky of Appro stressed that exascale computing will rely heavily on collaborative efforts. No one company will achieve [exascale] alone, said Kenisky.
Despite the buzz around pursuing exascale, the panelists came down strongly in favor of the democratization of supercomputing access. When discussion turned to recent sentiments expressed by the Obama administration supporting pursuit of HPC development, one attendee asked whether the panel thought the government was doing enough to pursue achievement of exascale. Charles Wuischpard of Penguin Computing posed the question of whether or not it was worth the effort to pursue such a lofty goal when so few in the scientific community ever actually use the systems at the upper end of the HPC scale. What are we doing for the scientists not using these systems? Wuischpard asked.
On-demand HPC, in particular, allows for opportunities to organizations that might not have previously had the capital necessary to access supercomputers. For entrepreneurs and grassroots scientists, the computer is a tool; the supercomputer a very expensive tool, said Chuck Moore. I think we're really in a power-to-the-people age of scientific discovery and exploration. I'm particularly excited about the bottom end. I think we're going to see a lot of incredible things come out of that.