SC2011 One of the presentations I caught at SC11 was by GPU computing pioneer Ian Buck - which is a good name for a pioneer, I think.
Buck’s Stanford PhD thesis, Stream Computing on Graphics Hardware, capped his research into using GPUs as computing resources and his work to develop Brook, one of the earliest programming languages aimed at GPUs.
More ReadingWhy gov labs presenting HPC tech ≠ officials wolfing overpriced sushiStudent cluster warriors to face off on new battlefield: ISCSUPERCOMPUTER vs your computer in bang-for-buck battleStudent cluster warriors battle it out in HamburgNvidia suffers as TSMC's 28nm ramp steeper than expected
This effort, of course, caught NVIDIA’s attention. They brought Buck aboard six years ago; he founded NVIDIA’s CUDA team and the rest is sort of history. History that he laid out in his talk at SC11 (watch the video here). He takes us from the earliest days (2002-03) of using GPUs as accelerators to where it is today. And, by the numbers, they’ve come a long way.
Right now there are more than 350 million compute-capable GPUs in the world: more than one for every man, woman and child in the US. Or one for roughly every 20 people on Earth, assuming they don’t mind taking turns with it.
NVIDIA has served more than a million direct CUDA downloads and figures there are more than 120,000 CUDA developers. NVIDIA has also banked more than $100,000,000 of Tesla GPU product revenue in the last year, and it’s growing at a pretty good clip.
Things are bound to get more interesting as Intel starts talking up its >50 core MIC (Many Integrated Core) co-processor (Son of Son of Larrabee). Intel rolled out some silicon at the show, but it’s not expected to see volume production until the end of 2012 at the earliest.
AMD is also pushing its Fusion CPU+GPU architecture, which could be an interesting HPC choice for some workloads.
NVIDIA has had the accelerator field to themselves for quite a while now, but that may end in the near future. They’ll be rolling out Kepler in the coming months, which should provide a significant performance jump over current Fermi-class Tesla cards. ®