Author: Robert Hyatt
Date: 14:36:51 08/20/04
Go up one level in this thread
On August 20, 2004 at 16:47:52, Tom Kerrigan wrote: >On August 20, 2004 at 16:23:12, Robert Hyatt wrote: > >>>Ah, so, some qualification, in case I run the experiment and get different >>>results. How about you send me the version of Crafty that showed an improvement >>>from 512k to 1024k? >> >>No idea which version it was. I could try to figure out when I first bought my >>original quad xeon, and track the version back to that, but does it really >>matter? >> >>You ran a simple experiment and drew a conclusion that is as accurate as >>flipping a coin and calling heads. If you are happy with the coin toss model, >>fine by me. > >I just downloaded Crafty and ran bench on an Athlon 64 3000+ and an Athlon 64 >3200+ (the difference being the amount of L2 cache, 512k vs. 1MB) and the result >was _identical_. Looking at the log files, times were off by 0.01s here and >there, so the timing function seemed to have good resolution. IDENTICAL nodes >per second. > >I based my conclusion on one datapoint and a good knowledge of computer hardware >and chess programs. If you think my knowledge is as valuable as a coin toss, >fine. But a second datapoint (doing an experiment that you yourself suggested >would have a different outcome) only serves to support my conclusion. So, >Occam's razor, do I understand how computers work, or am I just really lucky? > >-Tom Again, I have no idea. I know what the results were a couple of years back. I posted them when someone asked the question "Is the 2mb L2 xeon worth the $5000 per processor price over the 512kb at $1000 and the 1mb at $2500?" I have made lots of changes to address the MOESI cache issues on the Opteron. It is certainly possible that these changes made larger L2 caches less important for recent versions. As I said, I don't know. But clearly testing 256K vs 512K doesn't provide much actual data to draw conclusions from. Obviously the 2048K chip was not 5x faster, so the cost was a major issue. I seem to remember that back then, going to 1024K was worth 10%, 2048K was another 7%. But clearly more than zero, just not enough to offset that ridiculous cost. Everything I have run on from AMD recently (opterons) have had 1mb L2. My dual 2.8 xeon has 512K. I don't have any absolutely identical machines with different cache sizes to test on to try to draw any conclusions. However, from lots of benchmarking by myself and by AMD, the recent changes have greatly improved performance on opterons, particularly the 4-way and beyond boxes, due to some significant changes to address the MOESI/NUMA issues. More I can't conclude without any way to do testing. I might look up the cache modeling software and try that to see what it says, for fun...
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.