Among the smorgasbord of BoFs, plenary sessions, benchmarking results and vendor news that came out of SC15 last month, a few topics stand out as being particularly noteworthy, including China’s increased TOP500 share, the rise in accelerated computing and the fledgling OpenHPC project.
HPCwire Senior Editor Tiffany Trader sat down with Intersect360 Research CEO Addison Snell at the show to discuss these highlights and more. Their conversation, recorded in the video below, begins with a recap of this year’s TOP500 results, which Snell characterizes as “an interesting look at but not a perfect proxy for actual system revenue share in the high-performance computing market.”
Snell doesn’t read too much into China’s upswing, or its converse, the drop in US list share, which is at its lowest point since the list’s inception. His stance factors in that the TOP500 does not represent the whole market as well as the fact that these additional Chinese systems, most of them, are not new. They are older machines that were benchmarked with help from their respective vendors.
The other big trend seen both in the TOP500 list and on the show floor this year was in the realm of accelerated computing. “The number of systems with accelerated elements, things like GPUs or Xeon Phis has grown now across one-hundred for the first time, from 90 to 109 on the list,” says Snell. “And we’ve seen other proof points in accelerated computing being reported by NVIDIA, being reported by OpenPOWER, and also Intel with its Xeon Phi processors. That seems like we are getting to a tipping point where we will be seeing a lot more with accelerators.” This observation fits with a recent Intersect360 Research report that found that 34 of the top 50 HPC applications are showing some level of GPU support or optimization.
Intel, however, is steering clear of the terms “accelerator” and “coprocessor” to better distinguish its forthcoming self-hosted Knights Landing Phi as a full-fledged processor. To this point, Snell remarks: “In the end, these are all processing elements of different sorts. And to make the distinction between what’s a processor and what’s an accelerator comes down to labels. In the current instantiation of the market, the public perception is that these are accelerators since they are attached to a microprocessor at some level. And although I take Intel’s point that Knights Landing can be configured as a native processor, in our current structure of research, we’ve been grouping it together with accelerators because that’s the way that the market perceives them. I think with Knights Landing that might start to change and we’ll talk about these things in a different way.”
The discussion of Knights Landing kicks off further commentary about the swing toward specialization away from standardization, the difficulties of porting code in this new era, and what Snell refers to as “one of the most significant announcements,” the OpenHPC project that was announced by the Linux Foundation a few days before SC. (Watch the video for details.)
Snell’s overarching view of HPC is refreshingly optimistic in an environment that can sometimes be laser-focused on solving the next challenge, such that the big picture is lost. He observes that the great thing about high-performance computing or supercomputing is that science, being more or less infinite and requiring an unending level of complexity, will drive the need for more and more powerful tools in order to reach new levels of innovation. He adds: “until we wake up and decide that we’ve reached the end of science, and we say ‘okay science is done, we figured it all out, now we can stop,’ supercomputing will continue to go forward, the question is under what model this will take place.”