Author: Gian-Carlo Pascutto
Date: 15:05:48 05/02/04
Go up one level in this thread
On May 01, 2004 at 06:21:49, Tom Kerrigan wrote: >I got access to a quad proc computer this evening and timed how long it took to >search the BK positions to 10 ply, varying the # of threads. Here's the result. > >threads time(s) NPS(k) time NPS search efficiency >1 943 206 - - 100% >2 585 401 1.61x (as fast) 1.94x 82% >3 414 569 2.27x 2.76x 82% >4 365 742 2.58x 3.60x 71% > >I was wondering how this compares to other programs. My program uses a simple >implementation of YBW, which I found to perform significantly better than >ABDADA. (The latter seems to perform fine until you turn on null move.) > >I recall most algorithms in the ICCA Journal claiming excellent scaling for >4-way, usually over 3x IIRC. So I'm a little disappointed by 2.58x but I wonder >how other people measure speedups. Is it with null move on? Also, the set of >positions makes a huge difference. On 11 of the BK positions I get > 3x but the >overall average is dragged down by a few < 2x positions that take a long time >to search. The numbers for early Deep Sjeng were 1.7 on a 2x, and 2.65 on a 4x (with a standard error of up to 0.2...they are very variable). So this sounds very reasonable for a start. I used Hyatt's Quad Xeon. It's not always reliable to compare to other people because of the high variance and different testing methods. Don't stare blind on the ICCAJ numbers, just read how those numbers were (supposedly) achieved and you will understand why. You can run for example Crafty on the same positions and do your own math. That'll give you a baseline performance you *can* directly compare to. -- GCP
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.