During ISC’10 NVIDIA hosted a special session of briefings that our man Rich attended and wrote about. NVIDIA has put Andy Keane’s slides from that session online, and I was just trolling through it. Here are a few of the items that I noted
In 2008 NVIDIA’s R&D budget was close to $900M — by way of comparison, Cray spent $51.77M on R&D in the same year, and SGI spent $13.8M. I know: NVIDIA is still mostly a consumer products company, so the comparisons aren’t great, but still. Intel and AMD spent $5.722B and $1.848B, respectively, in the same year.
7 releases in 3 years — a slide in Andy’s deck highlights the fact that CUDA (and its associated tools) have had 7 releases in 3 years. The context is that this is a good thing, and it probably is, to a point. But NVIDIA cannot continue that pace and still grow the adoption base. At some point the churn in the tools drives away developers.
Community code ports — there is a fairly large body of community science codes that are now ported to CUDA. The deck shows about 25, with apps from WRF to AMBER.
5 PFLOPS — Andy showed the slide that NVIDIA has been talking to recently that shows a notional pure x86-based 5 PLFOPS machine at 20MW, and then a GPU-enabled hybrid machine at 10MW for the same theoretical peak. And we are shooting for an EFLOPS in 20MW by decade end?