Author: Robert Hyatt
Date: 07:04:28 08/24/04
Go up one level in this thread
On August 24, 2004 at 01:11:49, Tom Kerrigan wrote: >On August 23, 2004 at 23:26:28, Robert Hyatt wrote: > >>Of course it is _always_ easy to either criticize or guess. I tried to offer >>_real_ data. This little program will allow a lot of tuning, although I wanted >> >>the data. I came up with some numbers that at _least_ are based on some sort of >>reasonable approach to measuring cache usage. And it is a third data point that >>directly supports my 512K/1024K/2048K data and Eugene's 1.5m/3.0m data. We both >>show speed improvements as cache gets bigger, beyond 512kb. So did the >>cachegrind program above. three out of four showing improvement is fairly >>convincing. one out of four seems to be an anomoly of some sort, as yet >>unexplained. > >Yes, yes, real data this, real data that. Get off your damn "real data" high >horse. > >Are you saying that Anandtech's measurements are not real data? I am saying I have no idea what they are reporting. I haven't read it and don't really care... > You point out >that they messed up compiling TSCP for one of their articles. You ask for their >exact test setup. But this is just cheap psychological BS to try to discredit >the data in peoples' minds without actually saying anything about the data. Hey, >if you want to play scientist (you are supposedly a computer "science" >professor, right?), how about you try to reproduce their experiment? Get a >Sempron and an Athlon 64 and take the measurements. That's what a scientist >would do, scientific method and peer review and all that. Until you do that, you >can't legimiately dispute their data. > I don't believe I have disputed _any_ data. I disputed _your_ "proof" because it was not a proof of any kind... Back to the real world... >Besides, I took my own measurements, remember? Between a 2GHz 512k L2 Athlon 64 >and a 2GHz Opteron. They performed EXACTLY the same. How is that not real data? >So you're basically calling me a liar, with absolutely no justification. Well, >until you run the same experiment and get different numbers, I say you can shut >the hell up. Again, for the last time, what I said was this: "running crafty on an Xkb L2 cache processor, then running in on a 2Xkb L2 cache processor, getting the same search speed each time, does _not_ prove that the working set is < than Xkb." I have three sets of data. 1. intel xeon pentium II, 512KB, 1024KB, and 2048KB of L2 cache. Going from 512K to 1024K was 10% faster. Going from 1024K to 2048K was 7% faster. 2. Intel IA64. Going from 1.5MB to 3.0MB was 10% faster although the caches were not identical since Intel changed the associativity. 3. The data I produced here yesterday which also showed that cache misses dropped way down for caches > 512K total. Now I think I will "shut the hell up" since the discussions with you never go anywhere useful. You might try doing the same thing yourself... > >Speaking of the scientific method, it's interesting that nobody could reproduce >your Xeon experiment if they WANTED to, because you have no freakin' clue what >version of Crafty you used. It's not that hard--just try to remember vaguely >when you did the experiment and then figure out which version of Crafty you were >on at the time. I really do not understand the problem. The only reason why you >haven't done this already that I can think of is that your experiment was bogus >and you don't want people figuring that out. Based on yesterday's results they could use current source, wouldn't you think? Why would I run a bogus experiment to help me decide on whether to spend $1000, $2500 or $5000 per processor? I ran the test, and asked the question is 10% worth another $6000 dollars in a box that was going to cost about $10,000 with 512K L2 processors? Was another 7% worth _another_ $10,000?? The answer to both was "no". Had it been 50% things might have been different. I didn't run that test for _you_. I ran it for _me_. I wanted a real answer to make a financial decision. You can take your "bogus experiment" stuff and drop dead for all I care. It took a good bit of time to get a company to help me run the test. I didn't do it "lightly". Unlike most of your tests... > >-Tom
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.