Author: Dann Corbit
Date: 17:14:46 05/14/01
Go up one level in this thread
On May 14, 2001 at 19:50:25, Dieter Buerssner wrote:
>On May 14, 2001 at 17:27:15, Dann Corbit wrote:
>
>>You should be aware that having an 8 CPU system will be incredibly expensive,
>>and it won't blast everyone out of the water. Double CPU speed and you gain 50
>>ELO. So if you can run 7 times faster, you have added 350 ELO, which is a lot
>>but not insurmountable.
>
>I think, for your estimtion of added playing strength, you should do at least
>log2(number_of_cpus)*50 instead of (number_of_cpus-1)*50 Elo ...
I am wondering about this huge reduction. SMP loss is not nearly so great. I
think also it makes possible a earlier (or conceivably later) discovery of a key
position.
I have not seen any indication that ELO gain drops off with exponential gain in
speed. Perhaps there are studies of which I am not aware that indicate
exponential increase in strength is sublinear in ELO improvement.
Consider an 8 CPU system, with each CPU at 1GHz. We obviously won't get 8Ghz
throughput, but we won't lose all that much either, because with interlocked
exchange {which is all you really need for a chess game} contention will
{should} be low.
For this reason, I estimate the real loss will be only 1CPU out of 8 if
programmed properly. By the logarithmic calculation, a 64 CPU system would
provide only 6x performance, which means a 10% utilization. That seems absurdly
low to me.
>You are obviously aware, that other things will factor in, that may make this
>estimation optimistic nevertheless. However, I agree, that for a SMP capable
>program, 8 CPUs will add a lot of strenght.
But only 3x faster than a single CPU system? You certainly must be assuming
that there is a severe drop-off in ELO with added compute power.
Generally quoted we see 40-80 ELO estimates for each doubling in CPU speed.
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.