Author: Robert Hyatt
Date: 08:03:06 01/16/03
Go up one level in this thread
On January 16, 2003 at 03:13:11, Aaron Gordon wrote: >On January 15, 2003 at 21:30:56, Robert Hyatt wrote: > >>On January 15, 2003 at 19:30:14, Aaron Gordon wrote: >> >>>Here's a little bit of math to prove my point. This cpu speedup factor was from >>>Hyatt, by the way. The P3-1GHz vs P4-1.6GHz was tested here. Anyway, as I said >>>before the 8x1GHz P3 would be > P4-2.8x2 w/ SMT. Here's how. If a P3-1GHz = >>>P4-1.6GHz you can figure this... >>> >>>Lets take the total MHz into account. 8x1000 * 8.0 p3 GHz, * 1.6 = 12.8 P4 GHz. >>>Getting a speedup of 5.9 (1+(8-1)*0.7) you get 5.9 p3 GHz / 9.44 P4 GHz. >>>For the P4.. (1+(2-1)*.7) = 1.7x speedup, also add 30% for SMT. This comes out >>>to 2.21x speedup. Take 2.8 * 2.21 and you get 6.188 P4 GHz. >>>9.44GHz > 6.188GHz P4 GHz. 8x1GHz P3 would be over 52% faster. >> >> >>Except you overlook the PIII vs PIV xeon numbers I posted. My dual 2.8 is >>almost exactly twice as fast as my quad 700. The quad is PIII xeon, the >>dual is PIV xeon with SMT on. > >I was going on my numbers. So far the binaries I've compiled have been the >fastest I've seen so far so I bench with those, profiling for each individual >chip. P4 Binary for the P4, P3 binary for the P4, etc. P4-1.6GHz gets 620kn/s >and a P3-1GHz gets about 600kn/s. > >I never saw a post with both being compared. I do remember some posts a while >back.. you said something about 1.2mn/s in Crafty with GCC and 1.6mn/s with the >IntelC compiler on the Quad. As far as the P4-2.8 w/ SMT benchmarks go all I've >seen you post ~1.9-2.0mnps runs. This is a far cry from 3.2mn/s (double the quad >700). A dual 2400+ MP gets almost 2.2 million nps in the Crafty v18.11 >benchmark, btw. IntelC 5.01 compile w/ profiling. The PIV xeon dual seems to be running about 2.5M nodes per second in typical positions. For some positions it is slower, for others it is faster. Same for the quad. I'll give you a one-cpu test on the quad, and a one-cpu test on my dual. You can extrapolate from that... Actually I can't do that. The one-cpu test on my dual needs to use SMT to get a true reflection of the speed of one processor, but I can't limit two threads to one physical cpu. And I don't have a CPU terminator so that I can remove one cpu and run that way either. But as I said, general testing has the dual at roughly 2x the quad, by any measure I try. NPS. Time to solution. Etc. Since both run four threads, even the overhead is similar.
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.