Author: Robert Hyatt
Date: 09:39:05 09/28/02
Go up one level in this thread
On September 28, 2002 at 04:28:02, Aaron Gordon wrote: >On September 27, 2002 at 23:42:03, Robert Hyatt wrote: > >>I didn't run the SMP tests for AMD, I don't have a one here and have no plans >>to get one. I posted a chart of data others provided. I don't even remember >>which position we used now. All that was significant was that all the speedup >>numbers (raw nps, not parallel search times) were in the 1.4-1.5 range with >>AMD, and 1.8 and above for the intel boxes... >> >>I personally believe it highlights a memory bottleneck... > >I don't think it's fair for you to find the slowest possible binary for the AMD >and some IntelC5 binary and then claim that the speedup is slow. I don't think >it's fair either if someone takes a slow binary for a P4 and compares it to a >fast binary for an AMD cpu. I'm not doing that. But you are missing the point as we are not comparing speeds between AMD and Intel. I run _any_ executable on AMD using one cpu, then the same using two, and compute the NPS speedup. I do the same for Intel. It won't matter whether the executable is fast or slow, as I am not comparing nps between intel and AMD. I am comparing the NPS speedup from 1-2 cpus on AMD against the NPS speedup from 1-2 cpus in Intel... Fast or low executables won't make any difference in the _ratio_ I was looking at. Slow executable on AMD will still see a proportional speedup. Because the raw NPS is not important, the ration of 1 cpu to 2 cpus is all that counts here.. And AMD has problems... Not major problems, but problems nontheless... > >You seem to conveniently forget the benchmarks I've done and other people here >have done. Take a look at my latest graph of crafty results: >http://speedycpu.dyndns.org/crafty/craftybench4.jpg >Note: the P4 2.76GHz is an overclocked 1.8A northwood at 153.5fsb(614MHz RDRAM). I'm not forgetting _anything_. Benchmark nps does not matter whatsoever to _this_ discussion. It is _only_ the ratio of 2 cpu time to 1 cpu time for each specific processor. It shows that it is harder to run two cpus wide open on AMD than it is on Intel. > >Now, the SMP binaries I have are able to produce a 1.7x speedup in the >benchmark. You claim the P4's get 1.8x, thats fine. Take the P4-2.76's result >(1,120,011 nps) and multiply it by 1.8. You get 2,016,019.8 nps. Not too shabby, >right? Well.. take the 1.86Ghz XP and multiply it's nps by 1.7 and you get >2,035,330.1. Still faster. Now, if you're saying, "Well yadda yadda is >overclocked and etc etc". Yeah, and even faster things will be released here >shortly. I can guarantee the P4-2.76 w/ 614MHz RDRAM would be as fast or a hair >faster than a standard P4-2.8. The AthlonXP at 1.86 would be more around a 2300+ >if such a thing existed. Again, you are missing the point. I didn't say AMD was _slower_ than Intel anywhere. I simply said their two cpu machine does _not_ scale as well as the Intel duals. Nothing more, nothing less. That remains an easy to prove fact... > >Moving on to the future.. P4-3GHz will soon be released as well as the 2800+ >(being announced on October 1st). Lets do some rough guessing. If a P4 gets >1,120,011 nps @ 2.76 it should get about 1,217,403 nps at 3GHz and thats >probably still having the RDRAM clocked to insanity. Take the 2.52GHz AthlonXP @ >1,578,197. At 2133MHz (AthlonXP 2600+) it should do about 1,335,831 nps. Again >do 1,335,831 * 1.7 and 1,217,403 * 1.8 and you get: >2,270,912.7 nps for the dual XP 2600+ (2.13ghz) >2,191,325.4 nps for the dual P4-3GHz. Maybe or maybe not. But it _still_ doesn't change the fact that the dual AMD is less efficient (should optimally be 2x faster than a single) than a dual ]intel... > >Since Crafty is pretty linear you know these numbers are very close to the >actual results. So far from what I've seen Pentium4's need an entire GHz more >and twice the L2 cache just to come close. This is what I call a $500 keychain.
This page took 0.01 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.