Author: Robert Hyatt
Date: 07:02:23 02/21/03
Go up one level in this thread
On February 21, 2003 at 05:04:29, Matt Taylor wrote: >On February 21, 2003 at 00:49:18, Robert Hyatt wrote: > >>On February 20, 2003 at 20:10:58, Eugene Nalimov wrote: >> >>>On February 20, 2003 at 19:09:18, Tom Kerrigan wrote: >>> >>>>It seems you know a lot more than I do about instruction latencies. How do you >>>>explain Crafty's disproportionate speedup on IA-64, then? And why would you >>>>think Crafty's performance is a good predictor of other chess programs on IA-64 >>>>when Crafty is so much different from many other programs? >>>> >>>>Of course, all we have is Hyatt's word that Crafty does well on IA-64, although >>>>he's never seen it in person, and people who have seen Crafty run on IA-64 in >>>>person seem to contradict him... >>>> >>>>-Tom >>> >>>I saw Crafty on Itanium2 in person, and results were closer to what Bob reported >>>here than to what was reported here recently. >>> >>>Thanks, >>>Eugene >> >> >>I have only seen _one_ bad Itanium2 report. I have seen _two_ good reports, >>one from Eugene, one from someone inside Intel. I have a problem with the >>Intel version because of the spec stuff I did and a NDA that is in force. >> >>However, there are other good 64bit results from the alpha 264 chip as well, >>factoring 800K from 600mhz would reach beyond 1600K at 1.25ghz, which is the >>fastest 264 I have seen. >> >>Eugene has never told me what he did to produce good results. Nor do I know >>what Intel did specifically as I was not really involved and it was just a >>"heads up" from someone there about an interesting result. Whether either >>(Eugene or Intel) tweaked the code I don't know. I specifically know that Tim >>did not for the 600mhz 264 results I posted here, he compiled that code as is >>and ran it, without even testing to see if compact-attacks was better than the >>default or not. >> >>I don't understande all the 64 bit nonsense myself. It seems pretty obvious >>that for 64 bit operations, they are going to be faster, _period_. For 32 bit >>programs, who cares? Sort of like benchmarking a non-vector program on a Cray. >>Of course it will not run very well. > >Eugene reported 3M nps for single Itanium, right? No. NO one has reported > 1.6M for 1ghz parts. But 1.6M for 1ghz is pretty good, and it as as fast or faster than any single-cpu NPS I have seen, even better than the 21264 by a small margin. > >IIRC, my own results posted a while ago were about 1.6M nps in Crafty on a dual >1.67 GHz AthlonMP, so 800K nps at 1.67 GHz single-CPU. At 2 GHz, Athlon would be >roughly 1M nps. Itanium is faster then by a factor of 3 at half the clock speed. >Half the clock speed means it has equivalent theoretical IPS (6x1 GHz vs. 3x2 >GHz). Assuming -all- ops were 64-bit and therefore Crafty derived a speedup of >2, that does not account for the additional gain which is rather significant. > >Yes, Crafty will run faster on a 64-bit machine, but will it run that much >faster? Unless I have made some error, I don't see how the 64-bit machine word >could solely explain the speed gains on Itanium. I don't really know where the >extra speed came from. Maybe it's from the cache. Maybe Itanium comes much >closer to 6 IPC than Athlon does to 3 IPC. There are any number of explanations, >but I don't know because I have no idea how efficient IA-64 compilers are. Me either for the compiler question. But the 64 bit stuff is very important. All my move generation and evaluation (and lots of other stuff) depends on 64 bit operations, which translates into pairs of instructions on an X86 and a single instruction (actually a single part of a VLIW instruction) on the IA64... I've always assumed that was the difference, since I have seen several different results comparing 32 to 64 bit machines, and in every case, crafty speeded up more than the non-64bit chess engines... > >-Matt
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.