Author: Dann Corbit
Date: 14:28:36 02/17/04
Go up one level in this thread
On February 17, 2004 at 16:49:13, Bob Durrett wrote: >On February 17, 2004 at 16:14:03, Dann Corbit wrote: > >>On February 17, 2004 at 15:59:48, Bob Durrett wrote: >>[snip] >>>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! >> >>NPS as a measure of chess strength is useless. >> >>I can make an evaluation that just counts the material on the board and does one >>million NPS. >> >>A 200K NPS smart evaluation will clobberize it. >> >>Forget NPS as a measure of chess strength. The only usefulness of NPS is to >>measure a chess program against itself. People don't even have the same meaning >>for the term from program to program. >> >>So if I run a program on hardware X and get 500K NPS and I run the same program >>on hardware Y and get 1500 NPS, then I know that the program runs 3x as fast on >>hardware Y. That is all. I don't know anything about how strong it is in >>relation to another program or even if it will play a lot better. >> >>I only know how much faster it runs, relative to itself. > >That suggests that we need to identify a more useful measure of performance. And that depends on what you want to do with it. Do you want to solve tactical positions? Then we can measure with EPD test suites. Do you want to measure strength of play? Then things like the SSDF measurements can be used. >Once a useful measure is found, it remains to evaluate value of various >candidate improvements and express that value in terms of the chosen measures. Aye, there's the rub. >So maybe a better approach would be to identify the most time-consuming >processes in a conventional chess engine and then "hardware-ize" those >processes. If you are looking for a revolution in chess, that isn't it. It is orders of magnitude harder to change silicon than machine instructions. >Somehow, the expected benefits would have to be expressed >numerically before committing to the time, effort, and money required. That can be done simply enough. >How to measure this? I hope we are not falling back to SSDF! Probably, that is the best way to measure. But we can also do some extrapolations based on profile data. If you are looking for rapid improvements in algorithms, then hardware is the last place to look. To compile, link and run a change in software is 15 minutes tops. To make the same change in hardware is 15 days, minimum.
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.