Author: Vincent Diepeveen
Date: 05:45:58 05/27/05
Go up one level in this thread
On May 26, 2005 at 16:29:46, Joshua Shriver wrote: >Interesting... so how many of the 32 cpu's are CPU vs FPGA custom chips? It has 32 custom FPGA cards. >Have there been any stats as to nps on each fpga, and what kind of bus does it You are interested in objectivity, the hydra team is a very bad person to ask in that case. They are interested in blowing it up to unheard dimensions. >rest on? I thought the FPGA's where attached on a PCI backplane, and if they're >working with that many, have they come across any latency problems over the PCI >bus. That's why they do not have hashtables inside the fpga cards. >Hydra intrigues me, since parallel programming is what inspired me to want to do >an engine. (Besides also loving the game) Their parallel algorithm isn't worth much in the real world i guess. I don't give a penny for their parallel efforts. If they would have a good parallel algorithm, hydra would already run at 1024 cpu's by now for sure. Also the way it gets shown to the world is kind of the usual way. First dumb down 1 cpu a lot by not using hashtable last 6 ply, then show a good speedup of a lobotomized 1 cpu versus 32. We know that drill from the past already. Like cilkchess (MiT), a fast bitboard engine, getting just like 2500-5000 nodes a second at a single cpu at 512 cpu's 500Mhz, and diep a program 20 times faster, gets at the same hardware at 512 cpu's 10000-20000 nodes a second a cpu. I mean, cilkchess team can claim whatever speedup they have, but without cilk their program at 1 such cpu gets 200000 nodes a second, as it's a simplistic bitboards engine with near to no eval (crafty has 4 times more knowledge in eval than cilkchess in fact). So from my viewpoint they are missing a factor 40+ somewhere in the compare to the single cpu. See the problem of the parallel speedup of Hydra? They can brag whatever about speedup, as long as they aren't comparing an optimized single cpu version with the 32 processor version, it's not a very fair compare. You can get a million nodes, but if all what you do with it is search 6 ply in a highly selective manner, that still is 6 thin plies. It isn't 7 ply which i get with perhaps at most 100000 nodes, without any forward pruning (just nullmove). Then we didn't discuss the huge evaluation difference even. Rudolf Huber (SOS) calls that: "they first slow it down, or make the branching factor horrible, in order to be able to claim a better speedup". >-Josh > >>It's the fact that it's got 32 cpu's and the fact they can daily test at that >>that makes it so interesting. >> >>A node in hardware is like water in the sea. Majority of them are useless to >>you. >>
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.