Author: Vincent Diepeveen
Date: 04:20:20 08/18/00
Go up one level in this thread
On August 18, 2000 at 02:11:40, Dann Corbit wrote: >On August 17, 2000 at 23:43:57, Vincent Diepeveen wrote: >[snip] >>>Assuming a doubling of CPU power every year, how long until Diep can achieve 18 >>>plies on average at 40/2? >> >>Oh well in wcc 99 at 3 minutes a move at a quad xeon from Bob, >>with 400mb hash in endgame i searched always 15 to 20 ply. >> >>However endgame of DIEP was real BAD then. Now after i improved endgame >>considerable, it suddenly searches a lot less there! i haven't figured >>out yet why it doesn't search that deep in endgame anymore. >> >>My focus obviously is not at depth, but at evaluation! >>That is, first get a few plies, then fix evaluation! >> >>At 5:43 AM here it's hard to calculate, taking into account bigger >>hashtables as i have now! >> >>First let's do a few experiments next few weeks with this 256mb RAM dual 800, >>then after that let's extrapolate it! > >We can do the same calculation for Crafty fairly easily. > >The CAP database has an average of 13 plies at 12 minutes of CPU time. It is >predominated by opening positions, but we'll call it good enough as an average. > >That is 12 minutes of PII 300 MHz to reach that depth. If we assume an average >branching factor of 3, that would give 36 minutes to 14 ply, 108 minutes to 15 >ply, 324 minutes to 16 ply, 972 minutes to 17 ply and 2916 minutes to 18 ply. > >Now, this is only a PII 300. If we up that to 900 MHz machine we are back to >972 minutes. Obviously, you could do a lot better on an 8 CPU machine or >something exotic like that, but we will stick to "state of the art and yet not >out of reach" machines of today. > >40 moves in 2 hours averages 3 minutes per move. >972/3 = 324 times faster. >That would mean in 9 years, by hardware advances alone, Crafty should make about >18 plies on average at tournament time controls on a single CPU, fairly >state-of-the-art machine. > >On the other hand, it seems that a 64 CPU Alpha might be right about that >neighborhood in power (give or take a factor of 2). So if you have a fat >wallet, maybe nine years from now is tomorrow. First of all a 64 cpu machine is like, let's guess: 50 million dollars? So also in 9 years of time we can't afford that. Secondly a 64 processor alpha is perhaps not having shared memory, so getting a good speedup is real tough then. But i think bigger hashtables are giving your goal quicker as you think.
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.