Author: Ratko V Tomic
Date: 20:10:19 11/19/99
Go up one level in this thread
> to find the gain in elo per mhz you should focus on one program and note the >different elo ratings at different speeds . one thing i believe is that the >speed increase helps comp vs comp more than comp vs people . > There are only 3 platforms (A=P90, B=P200, C=K2-450) which cover all of the current programs (other than the dedicated units). But I don't think the elo/mhz approach would work, since these are different processors (different cache, diferent clocks/instruction, different memory bandwidth, etc). Instead one needs to find all the "bridging" programs between any 2 of the 3 platforms (AB, AC, BC)and obtain how the performance changes (statistically) between those two platforms. With several such programs between any pair of platforms, one will have a distribution of elo changes for A->B, B->C and A->C, thus one will have an average elo change and the spread for different programs (and their original elo uncertainties), giving the error margin for the estimate. For programs which haven't been tested on C, but were tested on A and B one would construct 2 ratings for C using the average elo jumps A->C and B->C, and also obtain the full error interval as the union of the A->C and B->C error intervals (or as the sqrt(Err(AC)^2+Err(BC)^2) ). I did something like that when the list came out, for the Tiger version they tested, and it came on top (as I recall, by a very little, within the error margins).
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.