Author: Robert Hyatt
Date: 08:12:46 07/21/00
Go up one level in this thread
On July 21, 2000 at 02:52:33, blass uri wrote: >On July 20, 2000 at 22:24:57, Robert Hyatt wrote: > >>On July 20, 2000 at 22:02:57, Chris Carson wrote: >> >>>. >> >> >>Nope. Please read again. I _clearly_ said it peaked at 1B nodes per second. >> >>The math is trivial: 480 chess processors, 1/2 at 2M nodes per second, >>the other half at 2.4M nodes per second. When you multiply that out, you >>get 1B+ nodes per second. Hsu figured that he _averaged_ 70% utilization >>for the processors. Down to 700M nodes per second. He then took 30% of >>that as his dissertation says that his parallel search efficiency was 30% >>for that many processors. IE a _honest_ 200M+ on average, while he was >>_really_ searching at 700M nodes per second but 70% of the nodes are search >>overhead. >> >>Deep Junior doesn't report search overhead. I will safely assume his is no >>better than mine, which means about 1/4 of his nodes are wasted on average, >>with peaks way above (and below) that number. > >I decided not to post opinions about the strength of deeper blue because we >cannot check facts about it and I will not convince you about it. > >I will post in this subject only when I think that the facts that can be checked >are wrong. > >In this case the effective nps of Deep Junior is 2.4M*3/4=1.8M >1000/2.4 is clearly bigger than 200/1.8 > >I also do not know if the 2.4M is not effective nps. >My Junior5.9 on pII450 does often 300Knodes per seconds. > >If Deep Junior is not significantly slower in nodes per second then I expect >pIII700*8 to be significantly more than 8 times faster in not effective nps. > >Uri Why? My NPS _drops_ in a SMP search. The PC doesn't have the necessary memory bandwidth to run at full speed. On my quad xeon, I lose 7% of my search speed if I run crafty on one cpu, and another compute-bound program on another cpu. If I run two other compute bound programs, I lose 15%. If I run 3, so that I am using one cpu and the other three are busy on something else, I lose 25% of my one-cpu NPS due to memory delays. an 8-way box is _much_ worse based on some testing on a Dell box I ran several months ago. And these are _raw_ NPS figures since I am just running a single cpu test, but contending for the tiny memory bandwidth limit the PC has...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.