At the SC10 Show in New Orleans on Wednesday morning I found myself as one of many conference attendees hustling down the uneven sidewalks on a surprisingly windy and cool day. Coffees in hand, we all managed to navigate the seemingly endless expanse of the convention center before the 8:30am start of the keynote session, featuring NVIDIA Chief Scientist Bill Dally’s insights on GPUs and the path to exascale.
Exascale supercomputers (those with roughly 1000x more computing capability than today’s fastest) will be critical to the advancement of science and industrial innovation this decade. But unless supercomputing architectures are designed more intelligently, these systems will require their own dedicated nuclear power plant to run them. Enter the GPU.
In order to reach exascale with a reasonable power budget a system must reach 50GFlops/Watt. Today GPU supercomputers can deliver 10% of this goal and CPU-only supercomputers just 1%. Assuming similar architectural and process advances occur for both GPUs and CPUs, GPUs will reach the exaflop goal by 2018. CPU systems will be only at 1/6th of the way to the exaflop finish line at that time.
Bill envisions that by 2018, GPU exascale systems could be housed in about 400 server cabinets (comparable to the largest supercomputers today) and will solve incredibly complex problems in computational research.
Check out Steve Keckler’s blog about this on the ACM.org site.
Or some press coverage of Bill’s SC’10 talk.