Author: Robert Hyatt
Date: 19:12:17 07/26/00
Go up one level in this thread
On July 26, 2000 at 15:27:36, Chris Carson wrote: >Just goes to show how much better the micro eval and search >algorithms are. :) TPR's speak for them selves. 96 DB >has fallen. :) Neither TPR for DB is very meaningful, since they cover 6 games. I could see 96 DB losing all 6 games to Kasparov, having a TPR of 2360 or whatever 1996Kasparov-400 turns out to be, and yet it would _still_ be better than today's micros. 6 game TPRs are meaningless. Particularly when it isn't a TPR at all, but a match vs _one_ opponent. For the same reason, the 1997 "TPR" is also meaningless when comparing to anything else. _unless_ you can find a TPR for another program vs Kasparov for 6 games. Those two numbers ought to compare and say something about ratings. > >IEEE micro should have been a source that a Ph.D in computer >science would quote in his/her first post, not a doctorial student >in psychology. :) I heard about this article today from another >person (different post), why did you not quote it? Did you not >know about it? You claim to be the all knowing source. :) I have mentioned it dozens of times here. We even had a long discussion about it when it hit the press a year+ ago... I have mentioned it to Ed several times over the past few days in fact... > >You are right, I should have read the article, my mistake. How >about you? I read it when it was published. In fact, I read it _before_ it was published. I referee for several different publications. > >Oh, try not to get to mad, but what position did HSU use to get >the NPS number, not that it matters for actual performance (we have >the TPR for that), but I would like to know? Last time you said there >is no position, is that still the case or is it in some journal article >somewhere thay you can do not know about (like you not knowing about >the IEEE micro article until today)? > >Best Regards, >Chris Carson There is no position. How about this bit of trivia: DB doesn't even count the nodes in the hardware. To quote Ken Thompson, who designed the Belle machine that became the core of the various chess processors Hsu built, "it takes as long to count a node as it does to _search_ a node. Why count 'em?" If I had his hardware, here is how _I_ would try to come up with a NPS figure: 1. how fast is each chess chip? (we have an answer for this) 2. How many chips will I have? (we have an answer for this too) 3. Can I drive the chips full-speed, or will it be something less. IE what is the duty-cycle for the chips on average? (we have an answer for this too). I could stop here and compute a RAW NPS. Or I could go on to: 4. If I use a one-processor search, and then compare that to a 2 processor search, then a 4 processor search, etc... all the way to a 480 processor search, what is the effective speedup. (we also know the answer to this as it is also available.) Stopping after 3, which reports numbers that are pretty similar to everyone else doing a parallel search, I'd get (assuming his numbers are right) about 700M, because his advertised "duty cycle" is 70% or so (and again, this is an _average_. It can go to 100%. And it can go to some lower bound that is probably close to 20%) If I wanted to be really cautious in the number I reported, I would factor in the efficiency of the parallel search compared to the serial search, and I would multiply that 700M by .3 and report 200M which is what he did. Or I could report 700M and _still_ be perfectly honest and comparable to the NPS other parallel searchers are reporting. If I _really_ wanted to know this number accurately, and I don't see why it is of anything other than passing interest since it has _nothing_ to do with the NPS of any other program, then I would design the hardware to maintain a node counter. I would then take the total nodes searched, by adding up the counters in all 480 chips, divide by the total search time, and post that as a raw NPS value. And I might scale that as in step 4 above, or I might not. It depended on what I wanted to use the NPS number for. But it clearly can't be compared to other programs in a meaningful way (other than it is obviously _way_ fast) any more than you can compare Hiarcs to Fritz in terms of NPS.
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.