Author: Keith Evans
Date: 13:17:58 02/17/04
Go up one level in this thread
On February 17, 2004 at 15:59:48, Bob Durrett wrote: >On February 17, 2004 at 15:36:30, Tord Romstad wrote: > >>On February 17, 2004 at 15:23:57, Bob Durrett wrote: >> >>>Bob, please indulge a "slow learner." I still don't get it. Are you saying >>>that the best way to get ***really*** high nps rates is with hardware [maybe >>>such as used by Hydra?] as opposed to using a PC? >>> >>>Incidentally, I am really feeling ignorant right now. How did Hydra get such >>>high nps? >>> >>>I hope you don't mind helping a beginner along on this confusing stuff. >> >>OK, I'll try: >> >>The CPU which sits inside your PC is, of course, not designed to >>play chess. It does not have any intructions to evaluate chess positions, >>generate legal moves, or any other chess-related tasks. When a chess >>program running on a PC performs such operations, each task is translated >>into a really big number of instructions for the CPU to execute. Executing >>all these instructions consumes a lot of clock cycles. >> >>Hydra, if I have understood correctly, uses hardware which is designed >>to play chess. It contains several processors which are built with the >>purpose of executing chess-specific tasks quickly and efficiently. >>Hydra's hardware probably *has* instructions for evaluating positions, >>generating moves, and similar tasks. Therefore, the processors don't >>have to execute nearly as many instructions for each node as the PC >>does. As a result of this, Hydra doesn't need as many clock cycles >>to process one node in the search tree, and this means that it can >>achieve a really high NPS despite a low clock frequency. >> >>Tord > >That is useful. Thanks. > >The way it is with true beginners is that for every answer they think of ten >more questions. Smarter people are smart enough to not show their ignorance, >whereas the beginners, like me, don't have anything to lose. : ) > >I guess by now you have figured out that I have another question! : ) > >It seems to me that there should be some way to provide a "cost" benefit for a >hardware feature, with the benefit measured in nps. > >For example, a "move generator chip" should be worth X nps. > >As another example, consider a "position evaluator chip." It is worth Y nps. > >This can go on for all the important functions performed in a chess engine. > >Would you care to estimate [guess at] the nps value of such chips? > >In this application, it would seem that a few well-designed chips might go a >long way! > >Bob D. You've got to be careful with NPS. One reason - not surprisingly Chrilly's hardware has a higher branching factor than typical chess software. So to search N plies deep his hardware has to search more nodes than software. Supposedly he limits the hardware search depth to 3 plies - this is probably picked to give the best overall performance. It will take significant PCI bandwidth to send new positions and retrieve results from the chess chips with just 3 ply searches in hardware. Reduce the depth and you increase the PCI bandwidth needed - increase the depth and you search too much. (Either searching junk or overlapping significant search with other chips.) If he just did the eval in hardware then it probably wouldn't make any sense, because it would take more time to transfer the position and retrieve the eval than it would just to do the eval in software. -K
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.