On Fri, Jul 18, 2008 at 4:49 AM, erik quanstrom <[EMAIL PROTECTED]> wrote: > very, very cool. thanks for the link, ron. > > i wonder if richard feynman (who is mentioned) might > have criticized the ops/s vs. time graph on p. 5 for being > overly fit to one end point -- the accounting machines > in the lower left? > > has anyone continued this graph to more recent times? >
Would not be hard. Cray 1 is 1e8 in 1978 and we're now at 1e15. It's not a uniform graph. Interesting that we hit 1e12 in 2000 or so and are now at 1e15 ...I built a 10T in 2002, so there is a slight slowing, but not much. But those Cray 1 vs. cluster FLOP numbers are different. A Cray 1 vector flop rate is not easily compared to a cluster flop rate. For $1m you can get a 40T system; the japanese earth simulator, 30T, probably cost about $500M to build including all costs. It's going to really ramp up with stuff like the Intel systems coming up. thanks ron
