Author: Robert Hyatt
Date: 08:24:15 05/06/04
Go up one level in this thread
On May 06, 2004 at 05:39:54, martin fierz wrote: >On May 05, 2004 at 18:24:11, Robert Hyatt wrote: > >[snip] > >>I see 16-cpu speedups vary from 6.0 to 15.3. That is not "very small >>variability" if we are talking about the same thing. > >we are not talking about the same thing because very clearly i was talking about >the 4CPU test. OK. My misunderstanding... I was talking about variability in general. In CB the variability is lower, because when I choose to split, I can look at the whole tree to see what is going on and choose to split wherever I want. In Crafty, I can split _here_ or wait until later, (here = current ply after at least one move has been searched serially). > >> The numbers are all over >>the chart, just like current parallel numbers are with Crafty... > >so be it - that's fine with me, because that will all become part of the >standard deviation and the standard error.... >in the end, perhaps you get a number which says > >speedup(4) = 3.1 +- 0.4 > >don't you realize that this is a much more valuable result than > >speedup(4) = 3.1 ?? Yes but I have never claimed speedup(4)=3.1 I have _always_ claimed that my speedup formula is a _rough approximation_ that varies even on the same problem set, or on different problem sets... IE it's a linear fit to a non-linear function... > >whatever it is, just post those numbers you just posted for DTS for crafty (and >vincent and GCP should post their numbers giving 2.8 and 3.0 too), then we can >calculate the stats, and see whether this is just a random noise thing or rather >a statistically significant effect. > >which would finally stop this fruitless discussion... I ran tests on the quad opteron. I believe I have logs for 1, 2, 3 and 4 cpus, I'll grab those off of a offline disk today and post the times as you requested... But I don't believe it will "stop this fruitless discussion" as I have posted the same kinds of numbers previously and it apparently didn't help at all. > >cheers > martin
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.