The always interesting Joe Landman had a great post yesterday at his blog, Scalability.org. In it Joe looks at SGI’s trends (read his post for that) and waxes practical about ClearSpeed.
My read on his analysis: ClearSpeed must change or die.
96 GFLOP at 12 W for new CSD.L part.
100 GFLOP at 100W for new nVidia part (according to Jim).
Great. Now look at cost ratios.
CSD.L: $5,000/96 GFLOP = $52.1/GFLOP
nVidia: $1,600/100 GFLOP= $16/GFLOP
After a bit of fairing up the cost comparisons by adding in the cooling and energy costs for the nVidia…it’s still half as much as the ClearSpeed solution. But wait, there’s more (bad) news
Now look at programming cost.
SDK for end users.
CSD.L: SDK cost ~$6,000
nVidia: SDK cost ~$0.00
…CUDA was a masterstroke. It lowered barriers to using accelerators right away. There are some valid criticisms of it (you should see some of the code we are playing with), but at the end of the day, it is possible for mere mortals to pull the SDK, compile code, and deliver applications at a very low relative cost.
I could go on with the quotes, but really the whole thing is worth quoting so go read it. Joe’s basic point is that the quality of the technology doesn’t matter: if the single stipulation that the solution is “good enough” can be met, cheap wins.
And this is absolutely true. We no longer have custom HPC hardware, even though it was arguably better than what we have now (the percentage of peak available to applications has gone down by an order of magnitude just over the course of my career). What we have now is much cheaper than what we used to have, and it’s good enough, which is why it’s ALL we have now.