Author: Robert Hyatt
Date: 20:28:12 06/29/00
Go up one level in this thread
On June 29, 2000 at 13:25:19, Tom Kerrigan wrote: >On June 29, 2000 at 03:38:38, Albert Silver wrote: > >>On June 29, 2000 at 02:21:10, Gregor Overney wrote: >> >>>>Not even close yet. That hardware would be approximately 1% of the power of >>>>the DB hardware. And that is being _very_ generous... >>> >>>1% is a pretty good estimate for a four processor machine using four P5-4 >>>running at 1.5 GHz using a four channel RDRAM bus that delivers 3.2 GB of data. >>>Estimate 500 kNodes per CPU times 4 = 2M Nodes = 1% of DB's avarage performance. >> >>If one CPU achieves 500k nodes, I doubt very much that 4 CPUs will achieve 2M >>nodes, unless 100% efficiency has been achieved. Crafty is apparently the most >>efficient at this level though only Bob would be able to say how well it should >>do. > >No, believe it or not, Bob Hyatt is not the only competent chess programmer in >the world. > >In any case, who says that DB was searching at 100% "efficiency"? (I assume you >mean overhead from the parallel search.) DB was composed of HUNDREDS of >processors, and each processor was pretty localized from the others. Compared to >a PC program running on only 4 processors with a shared hash table, etc., the >overhead of DB must have been tremendous. > >-Tom THe numbers for DB are well-known. 500 processors running at 2M nodes per second per processor (2M on half, 2.4M on the rest). Roughly 1B nodes per second max. But typically it ran at around 700M. Hsu reported that his parallel search was about 30% efficient. Rather than report raw NPS at 700M, he reported "effective NPS" at 200M. My 1M nodes per second is really closer to 700K. But then I didn't deal with a message-passing architecture and two levels of parallel search.
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.