About 3 years ago, I was at a Sun HPC consortium meeting, where there was excitement over the possibility of getting 1TFLOP into 2.5 racks with an ultra-dense server. This was cool. It was awesome. One of the conference organizers was talking to me about this, saying it was the densest possible system (at that time).
Having just been through the accelerator card high level design process for a business plan/company concept we were pitching to VCs, I innocently (ok, well, not so innocently) asked …
“Well, what if you could get a real sustainable 1TFLOP in a 4U box?”
I was accounting for heat removal and other things. The person just looked at me, blinked rapidly for a moment, and then said “But you can’t do that!”.
Well, yes we could. Just couldn’t find anyone smart enough to fund the company to make it happen.
Why am I recounting this history? Many reasons. First and foremost, ATI and Nvidia have released 1 TFLOP chips. ATI’s is ~200 GFLOP in double precision from what I read, though I will find the right linky. Nvidia’s is ~100 GFLOP in double precision. So it shouldn’t be that hard, with sufficient number of GPUs, to create a unit with an aggregate 1TFLOP in 4U.
Other elements of interest: the worlds fastest computer is an accelerated compute. One of our readers points out to a new Clearspeed chip. I’ll pull my comments into another post on that.
All in all, the accelerated world is … accelerating.
Recently I noted that our market predictions for acceleration were pretty accurate. Actually now, in retrospect, we were wrong. So wrong.
We underpredicted the market.
Oh well … :)