Author: Robert Hyatt
Date: 08:11:18 08/24/05
Go up one level in this thread
On August 24, 2005 at 05:27:29, Daniel Mehrmannn wrote: >On August 24, 2005 at 03:33:37, Joachim Rang wrote: > >>On August 23, 2005 at 22:47:07, A. Cozzie wrote: >> >>>Bob let me run on his 8-way from AMD for an hour or so tonight, so I thought I'd >>>post the results. I think he is going to run Crafty on these positions as well. >>> This is only 3 positions so it doesn't represent anything statistically >>>significant, but it gives me a general idea of how I'm doing. >>> >>>Nolot 1: >>> >>>1. Ng4xh6 c4-c3 2. Nh6-f5 c3xb2 3. Qd1-g4 Bb7-c8 4. Nf5xg7 Re6-g6 5. Qg4xg6 >>>Qe8xg6 6. Rg3xg6 Bf8xg7 7. Ne4xd6 Bc8-d7 >>> = (-0.13) Depth: 14/40 00:03:27.46 889946kN (4290 KN/s, 28617 >>>splits, 1805 aborts) >>>1. Ng4xh6 c4-c3 2. Nh6-f5 c3xb2 3. Qd1-g4 Bb7-c8 4. Nf5xg7 Re6-g6 5. Qg4xg6 >>>Qe8xg6 6. Rg3xg6 Bf8xg7 7. Ne4xd6 Bc8-d7 >>> = (-0.13) Depth: 14/38 00:24:01.14 797115kN (553 KN/s, 0 splits, 0 >>>aborts) >>> >>>Larsen-Spassky: >>> >>>1... h5-h4 2. Be2xg4 Bf5xg4 3. h3xg4 h4xg3 4. Rh1-g1 Rh8-h2 5. Qc2-c3 Qe7-h4 6. >>>Ke1-d1 Rh2-h1 7. Rg1xh1 Qh4xh1 8. Kd1-c2 g3-g2 9. Nb1-a3 >>> = (2.57) Depth: 15/40 00:02:00.34 604796kN (5026 KN/s, 25748 >>>splits, 1329 aborts) >>>1... h5-h4 2. Be2xg4 Bf5xg4 3. h3xg4 h4xg3 4. Rh1-g1 Rh8-h2 5. Qc2-c3 Qe7-h4 6. >>>Ke1-d1 Rh2-h1 7. Rg1xh1 Qh4xh1 8. Kd1-c2 g3-g2 9. Nb1-a3 >>> = (2.57) Depth: 15/42 00:12:46.15 500809kN (654 KN/s, 0 splits, 0 >>>aborts) >>> >>>Nolot 10 >>> >>>1. Rf1xf7 Rf8xf7 2. Bd5xf7 Kg8xf7 3. Qd1-h5 Kf7-g8 4. Qh5-e8 Be7-f8 5. Nc3-d5 >>>Qb6xd4 6. Nd5-e7 Kg8-h8 7. Ra1-f1 Qd4-f6 8. Rf1xf6 g7xf6 9. Ne7xc8 >>> = (1.58) Depth: 14/42 00:02:20.42 715711kN (5097 KN/s, 26886 >>>splits, 1288 aborts) >>>1. Rf1xf7 Rf8xf7 2. Bd5xf7 Kg8xf7 3. Qd1-h5 Kf7-g8 4. Qh5-e8 Be7-f8 5. Nc3-d5 >>>Qb6xd4 6. Nd5-e7 Kg8-h8 7. Ra1-f1 Qd4-f6 8. Rf1xf6 g7xf6 9. Ne7xc8 >>> = (1.58) Depth: 14/39 00:09:09.82 373351kN (679 KN/s, 0 splits, 0 >>>aborts) >>> >>>Scaling: 7.75, 7.68, 7.51, for an average of ~7.6 >>>Speedup: 6.94, 6.37, 3.91, fo an average of 5 or so. >>> >>>The first two are more or less standard positions and Zappa gets a speedup of >>>6.5 The last is more or less the worst case as Zappa switches among mainlines >>>constantly and evidentially some real wasted work was done :( >>> >>>Of course in a game things work a bit better as you ahve filled hash tables, but >>>I'm still sort of disappointed. The biggest problem I have is that the worst >>>case is also the most important case: when you don't know what move to play :) >>> >>>anthony >[...] >> >>Thanks for sharing this. I think it is not too bad. On some positions you always >>will have a bad speedup. But your data shows that generally your parallel search >>works okay with 8 procs. >[...] > >Not bad ?? This is extremly good and the best scaling i've ever seen so far. >Looks like anthony did a perfect job. > >Best, >Daniel I'm not sure what you mean by "best scaling so far"... The NPS numbers don't mean much and we both get good results there. It is possible to tune Crafty to get 8.0X nps scaling, but that results in a parallel search that is not as efficient in terms of speed-up...
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.