Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / processor coefficent for rybka
- - By Carl Bicknell (*****) Date 2008-01-16 23:35 Edited 2008-01-16 23:38
OK: this is an off-shoot from another post. I want to compile a list of processor efficiency scores. I want to do it like this:
The Core 2 architecture is our benchmark: it has a score of 1.00 for ONE core running in 64bit.

Someone has suggested that clock speed for clock speed the Pentium 4 has a relative efficiency of half that, i.e 0.5.
To get the TOTAL speed we would multiply the number of cores by the clockspeed (3Ghz or whatever) by a dual or quad processor efficiency i.e. 0.75 for 2 cores. Also rememeber to factor in 32/64 bit - the P4 was a 32bit processor.

So, for example, to compair a 2.8 Ghz Pentium 4 with a 2.33 Core 2 duo we have:
2.8 x 0.5 (coefficent) x 0.6 (32bit) x 1 (number of cores) = 0.84
2.33 x 1 (coefficent) x 1 (64bit) x 2 (number of cores) x 0.75 (speed loss from going to 1 to 2 cores) = 3.5

Thus to find the speedup 3.5 / 0.84 = 4.2 !

Amazingly a Core 2 duo running at 2.33Ghz is 4.2 times faster for Rybka than a Pentium 4 running at 2.8Ghz

Now then, please can people provide other coefficents for other processors?
So far we have:
Core 2 = 1
Pentium 4 = 0.5

Thanks! I will make a spreadsheed in the end and email it to anyone who wants it. We need Athlon 64, Phenom, Pentium 3, Pentium M etc etc
Parent - - By M ANSARI (*****) Date 2008-01-17 08:31
There are a lot of other factors involved ... even with same CPU's.  More cache on CPU die means much better performance per equivalent Ghz ... also more optimizations in the architecture of the CPU itself can make a difference.  Another thing is memory ... memory latency and memory speed have tremendous effect on performance of equivalent CPU's.  So to have an accurate estimate of performance you have to figure in a tremendous amount of different data ... how much cache on CPU, what FSB is the CPU running, what latency, which batch is the CPU etc ...

Performance seems to increasing at breakneck speed with hardware.  The new 45nm CPU's from Intel perform clock per clock about 10% better than the older 65nm CPU's (such as the Q6600's).  Forget P4 technology ... that is probably the worst performing CPU for chess and is not even worth writing about.  What would be interesting is the performance of the AMD Phenomn processors as they move to smaller die technology.  AMD still has better performance due to the on cpu memory controller when compared to Intel ... but that is not enough to offset the incredible clocks that Intel CPU's are able to get.  Intel is about 1.5 generations ahead of AMD with die technology ... and currently that is killing AMD.  This difference would seem to get even more profound as Intel also starts putting their memory controller on die.
Parent - By Carl Bicknell (*****) Date 2008-01-17 10:02
Well in that case we may need to specify like Core 2 quad @ 45nm etc with 4MB Cache etc and have a seperate coefficent for the different varients - but that's not hard to do.

I DO want to include older processors because several people have shown an interest and it is useful for compairing cpu's today with an older machine or for planning an upgrade to see how much improvement has been made.

I even have some coefficent scores of the old 1980's motorola 68000 processor.

The issue you raise about memory latency etc is obviously not possible to include so this guide will have to be an approximation - I can't improve on that. But the architectural issues of the CPU are exactly what the coefficent takes into account so this is not a problem.
Parent - By Vasik Rajlich (Silver) Date 2008-01-18 14:24
I would suggest going by NPS for this.

The coefficient for going from 1 to 2 cores is higher than 0.75, probably somewhere between 0.8 and 0.9.

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / processor coefficent for rybka

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill