Author: Robert Hyatt
Date: 08:02:03 02/18/04
Go up one level in this thread
On February 18, 2004 at 08:12:57, Vincent Diepeveen wrote: >On February 17, 2004 at 14:56:04, Robert Hyatt wrote: > >>On February 17, 2004 at 14:48:43, Gerd Isenberg wrote: >> >>>On February 17, 2004 at 14:02:01, Robert Hyatt wrote: >>> >>>>On February 17, 2004 at 13:48:16, Mark Young wrote: >>>> >>>>>Hydra >>>>>8 x 2.8 Xeon & 8 FPGA Cards >>>>> >>>>>Total games 47 >>>>>Stored results 47 >>>>>White 24 >>>>>Wins 19 >>>>>Draws 16 >>>>>Losses 12 >>>>> >>>>>Results 27/47 = 57.4% >>>>>Opponents Elo 2604 >>>>>N Opponents 31 >>>>> >>>>>Rating 2634 >>>>> >>>>>I see nothing in the results or games to suggest at this time. That Hydra is the >>>>>end all, be all future of computer chess. >>>>> >>>>>The last game I saw Hydra playing Shredder 8 on a P600 and down a pawn when >>>>>Hydra logged off. TC was 15 10. >>>>> >>>>>Dispite the hype... Hydra is not exactly blowing the top PC programs out of the >>>>>water even on much slower hardware. If the results are correct on playchess.com >>>>>for Hydra. >>>> >>>>You have to learn that hyperbole follows any new project. It is fast, and it >>>>plays pretty well. But it isn't _that_ good. IE It is not the second-coming of >>>>deep blue or some such nonsense. I did a few runs on an 8-way opteron two weeks >>>>ago and saw speeds that were around 16-17M nodes per second. Hydra's speed can >>>>already be reached by existing hardware, and without the issues of no hash in >>>>hardware, its speeds can be beaten with existing machines. >>>> >>>>IE the last version of Belle was first seen in 1980. By 1983 Cray Blitz was >>>>out-searching it. And we continued to get faster every year while Belle was >>>>stuck at the hardware speed it was built around. >>>> >>>>DB's 200M nodes per second average was incredible in 1997. I've searched 10% of >>>>that speed already and there are boxes around that will go maybe 4x faster than >>>>that. So today, it is possible to search about 1/2 as fast as deep blue, but >>>>with their hardware non-hashing problem, we might be faster today. >>>> >>>>Time marches on, hardware designs tend to be "flash in the pan" things, >>>>brilliant today, not so fast tomorrow, slow next year... >>> >>>but even FPGAs became faster and more sophisticated the next years. >>>It's probably a bit more flexible/portable architecture than former solutions. >>> >>>PCI(X)-bus is the bottleneck, so there is a tradeoff between PCI-bus >>>bandwith/latency and number of hashless FPGA-plies. May be it's even possible in >>>the future to do four(+quiescence) FPGA-plies with appropriate branching factor >>>and driving PCIX on the limit. Chrilly already implemented killer heuristics... >>> >>>I believe they will domitate the scene for a few years. >>>But of course i may be wrong ;-) >> >>I don't, myself. were they 20x faster than the next closest box, maybe. But >>right now they are _no_ faster and perhaps actually slower with the FPGA search >>being somewhat primitive. > >They search 500k-800k searches a second in software at 8 cards. Peek against me >was something like 900k searches a second in software. > >Each search in software is a 3 ply search nowadays. It's doing checks in qsearch >and using nullmove and it's not doing lazy evaluation (that doesn't work for >them like it doesn't work for majority of strong programs). I'd bet _anything_ it is doing lazy evaluation, because both belle and DT/DB did it. They called it a "fast evaluation" and a "slow evaluation". based on the clock cycles required to search a node (10) it looks _exactly_ like the previous hardware implementations, IMHO. Since Ken was involved, that probably cinches it. They might not call it "lazy eval" but I'd bet dollars to holes that they are doing the two-stage belle evaluation. > >You should therefore be able to figure out that the total nps is very very high. > >Do not compare this 90k nps with the world champs 2002. It did 2 ply in hardware >and 3 ply in endgame back then. > >Just like deep blue chips big forward pruning gets used in the hardware, >additionally nullmove here too. > >Each move it searches about 16-18 ply deep at 40 moves in 1:30 hour. That doesn't worry me at all. Look at the Crafty - Junior log from CCT6. Crafty searched 16-18 plies at 45 10 time controls...
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.