Author: Vasik Rajlich
Date: 03:53:46 05/29/04
Go up one level in this thread
On May 29, 2004 at 04:00:31, Gian-Carlo Pascutto wrote: >On May 28, 2004 at 21:27:16, Dann Corbit wrote: > >>I am guessing that the latest professional programs are not NUMA ready yet, >>though probably Deep Sjeng is. > >The Opteron doesn't really need NUMA code for 4 way, although it may start >to be an issue for 8 way. > >There are a few pitfalls, which Crafty was having issues with, but they are >not hard to fix, and the top pro's don't seem the have them. See my tests at >Graz with my machine there that I posted here - all of them scaled excellent. > >(including Deep Sjeng obviously :) > >>The faster the hardware, the better crafty does. I think (compared to others) >>crafty scales better. That would indicate better algorithms and inferior >>microoptimization. Also, the other deep programs do not seem to benefit nearly >>so much from multiple CPUs (though they may have possibly ironed this out). > >I have certainly seen no evidence for that - see my tests with them. > >>On the other hand, I suspect that if Deep Sjeng got the same 8 CPU machine, then >>it would win. > >I don't think so. The program still has weaknesses that a bit of >extra hardware will not overcome. > >>What sort of hardware are you bringing to the party? > >Uh, the provideed P4 2.8Ghz or maybe my Athlon64 2.2Ghz. > >>Now, Shredder Fritz and Junior might come on powerful hardware (e.g. 2.4 GHz AMD >>64 bit dual) but I am guessing that would translate to about 3 GHz throughput. >>[All the measurements I have seen showed them scaling very poorly] > >Again, I cannot confirm that at all! Also, both Junior and Fritz were >already 4 way last year. > >>Crafty -125 Elo (Crafty 19.12 and 19.13 are huge improvements on the >>predecessors) > >Crafty 18.12 is about 200 ELO worse than Shredder 8 - I wonder if >there has been that much improvement in Crafty. > >>So Crafty would have to overcome about 175 Elo to be evenly matched against >>Shredder in my crude model. So, with an 8-way box he would have about >>.65*2.4*8=12.5 GHz (let's say 12). That would be about 4 times faster than >>Shredder, Fritz or Junior. With +70 Elo per doubling, that would be 140 Elo. >>So crafty would have a real shot at it, even against Shredder. > >These maths seem rather flakey. First of all, 8 way does not mean 8 times >faster. A good estimate is a speedup of 6 on an 8 way machine, and quite >possibly less. Also, the SSDF has less than 70 ELO for a tripling of speed >nowadays, so 70 for a doubling is also way too high. > At this point nobody knows what happens on big hardware. We know that on some Athlon 1200s, 32 bit, tiny cache, with some book & autoplayer arrangement that both sides will complain about, Shredder scores about 75% against Crafty. What happens when you increase the hardware? Is it that Shredder selective search, which gives it a huge tactical advantage on weak machines, no longer provides any real benefits? Is it that Shredder's far better branching factor ensures that the better hardware, the more hardware advantage Crafty needs just to keep up? Maybe the diminishing returns are so great that evaluation starts to become the key point. Who has the better eval? Maybe Shredder's aggressively tuned eval works against shallow searches, but not against deep searches. We just need more data - nobody plays these types of games. There are also book questions. Shredder plays sharp with black. Perhaps it's good for the SSDF - but it led to two lost opening positions right in the opening (vs Fritz & Junior) at last year's WMCCC. Who really has the better book, and opening philosophy, for this particular situation? That's why we actually play the games ... Vas >Anyway, I'm happy that at least someone is reasoning about this. I think >if you redo the math with the adjusted estimations, and compute the chance >for Crafty to win the event, it'll most likely be 'statistically >insignificant' :) And so it will be for Sjeng. Then you can understand >where my 'claims' are coming from. > >-- >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.