Author: Gian-Carlo Pascutto
Date: 13:18:55 06/04/02
(This discussion started on the chess-engines list and got a bit too big, so I move it over) I started here: ----- Original Message ----- From: "Robert M. Hyatt" <hyatt@cis.uab.edu> To: <chess-engines@yahoogroups.com> Sent: Tuesday, June 04, 2002 6:48 PM Subject: RE: [chess-engines] Moves in first five ply > On Tue, 4 Jun 2002, Vincent Diepeveen wrote: > > > My datastructure and the one from Yace is a factor 2 faster > > than yours Bob at 32 bits processors. > > > > At 64 bits processors you win 33% or so in speed (see > > specbench for alpha processors. My guess that's however > > not the 64 bits so much as well as also partly because > > of doing 4 instructions a clock versus 3. > > > > So still a lot slower. > > > If you believe that, its ok by me. > > I don't... Isn't it possible to more or less objectively figure this out? We _have_ bechmarks of Crafty on Alpha and Intel/AMD. Best Alpha in SPEC2000: Compaq Computer Corp AlphaServer ES45 Model 68/1000 Alpha 21264C 1000Mhz Crafty ratio: 816 Average SPECINT over all apps: 679 Closest AMD matching machine in average SPECINT speed: Advanced Micro Devices Asus A7M266-D Motherboard AMD Athlon (TM) MP 2000+ Crafty ratio: 967 Average SPECINT over all apps: 662 If I interpret this correctly, this implies that for two machines which are within 2% in overall CPU speed, the 64-bits machine is SLOWER in running Crafty than the 32-bits one. Hmm. -- GCP And last posts were On Tue, 4 Jun 2002, Gian-Carlo Pascutto wrote: > ----- Original Message ----- > From: "Dann Corbit" <dcorbit@connx.com> > To: <chess-engines@yahoogroups.com> > Sent: Tuesday, June 04, 2002 8:20 PM > Subject: RE: [chess-engines] Moves in first five ply > > > > A profile of crafty shows that it is very even handed. In other words, > > there are no striking bottlenecks. In most programs, you can easily > > find some place and say that if you could speed up one or two places you > > would get a big boost. > > > > I suspect that what happens is that the bitboard stuff ceases being a > > bottleneck and something else becomes a more dominant speed trap. Of > > course, my speculation is even wilder because I have never profiled > > crafty on a 64 bit machine (despite that fact that we have many > > different kinds here). > > Considering Crafty basically consists of no more than fiddling around > with bitboards, I agree with Vincent that it would be RAM latency. > > That's a pretty nasty one, considering CPU's have been getting faster > faster than RAM has. Your hope would be that someday the internal > caches get big enough to contain all essential bitboard data. > > Already done. _none_ of crafty's data arrays used for updating the bitmaps adds up to a megabyte. IE the rotated lookup tables are 64 * 256 (=16K) entries of 8 bytes each. There are four of them. The remainder are much smaller and the entire kit and kaboodle fits in under 1 meg. Memory bandwidth is not a problem with any reasonable size of L2 cache. My xeons are 1M and work just fine. -- Robert Hyatt Computer and Information Sciences hyatt@cis.uab.edu University of Alabama at Birmingham (205) 934-2213 115A Campbell Hall, UAB Station (205) 934-5473 FAX Birmingham, AL 35294-1170 ----- Original Message ----- From: "Robert M. Hyatt" <hyatt@cis.uab.edu> To: <chess-engines@yahoogroups.com> Sent: Tuesday, June 04, 2002 9:54 PM Subject: Re: [chess-engines] Moves in first five ply > I don't follow the logic. For applications that are designed > around 32 bits, you "normalize" the two processors. Even though > in the case of Crafty, we have an application that is _designed_ > around 64 bits? > > mhz for mhz, 32 vs 64 is meaningless for applications that are > designed around 32 bits. There is no benefit. But for applications > that use the extra data density of 64 bit words, the advantage can > be significant. But then why are we not seeing a relative speedup on the Alpha? Crafty is optimized for 64 bits, yet when running on a 64 bits Alpha, it gets relatively SLOWER than the other apps, which are presumably not optimized for 64 bits. By your reasoning, Crafty should have a better relative performance than the average application, on the Alpha when compared to the x86 machines, but we see exactly the reverse in SPEC results. -- GCP ----- Original Message ----- From: "Robert M. Hyatt" <hyatt@cis.uab.edu> To: <chess-engines@yahoogroups.com> Sent: Tuesday, June 04, 2002 9:51 PM Subject: Re: [chess-engines] Moves in first five ply > How would you feel about a 1.5M nps crafty on a new IA64 at > 1.0 ghz. > > Not bad when no 32 bit processor can come even close to that at > more than double the clock speed... As I stated before, the raw performance of the machine is meaningless. If you'd put a 0x88 program on, it might very well also skyrocket in NPS. (Nice NPS anyway, IIRC the first generation IA64 was just pathetic with Crafty) -- GCP On Tue, 4 Jun 2002, Gian-Carlo Pascutto wrote: > ----- Original Message ----- > From: "Dann Corbit" <dcorbit@connx.com> > To: <chess-engines@yahoogroups.com> > Sent: Tuesday, June 04, 2002 7:44 PM > Subject: RE: [chess-engines] Moves in first five ply > > > > At some point, 64 bit CPU's are going to clobber the 32 bit ones [the 8 > > bit story and 16 bit story are sure to repeat themselves]. Right now, > > 64 bit chips are running at half the clock rate of the 32 bit chips. I > > don't imagine that they are going to suddenly release CPU's that have 4x > > the performance because it would kill off their current inventory of 32 > > bit stuff. I suspect we will see a slow dribble of technology updates > > (like always) unless something pushes them. > > The issue isn't whether 64-bit CPU's will get faster than 32-bit ones. > They will. The issue is whether (and how big) the benefit for a > (bitboard) chessprogram from being able to work with 64-bit quantities > is. > > The 32-bit CPU's didn't clobber the 16-bit ones so much because of > the possibility of using 32-bit data quantities over 16-bit quantities, > but because they were more advanced and faster designs. > > We're not debating whether the CPU's are faster or not - we're debating > whether the 32-bit to 64-bit transition in itself has significant > performance advantages. > > The original comparison pointed out that for two processors that are > of similar speed over a wide variety of benchmarks, the 64-bit one does > not have an advantage over the 32-bit one as far as Crafty goes. Here I simply disagree. "which" 32 bit processor will you suggest that can do 1.5M nps with crafty? Intel has a 64 bitter at 1.0ghz that will do this. And it will clock a good bit higher than 1.0ghz also, to further "spread the divide"... > > For me, that is a surprising result. It seems that Crafty does not > get faster when the bitboards fit in registers and can be manipulated > by single instructions. > Because you are using a processor that is clocked at twice the clock frequency? Why compare a 1ghz processor to a (nearly) 2ghz processor and conclude anything about efficiency there? Is there anything that suggests that the alpha is simply more "efficient"? To justify that clock frequency disparity? A machine twice as fast (clock freq) _should_ perform just as well as a 64 bit machine at 1/2 the frequency... Less would suggest that the 32 bit machine simply sucks badly. Mckinley certainly bears attention. It is producing some amazing numbers. > A possible explanation might be that, right now, the bitboards allow > the processors to take full advantage of their superscalar design by > always keeping a high amount of pipelines busy. When the bitboards > can be manipulated by single instructions, you suddenly can't issue > multiple isns per clock any more because of data dependencies. Your > secondary and ternary pipelines stay empty and you run no faster. > (The above paragraph is wild speculation) > That is the reason bitboards work well at all on 32 bit machines. It is difficult to keep 2 or more scalar pipes busy. pairs of 32 bit operations make this easier... and use cycles that would be otherwise lost... > -- Robert Hyatt Computer and Information Sciences hyatt@cis.uab.edu University of Alabama at Birmingham (205) 934-2213 115A Campbell Hall, UAB Station (205) 934-5473 FAX Birmingham, AL 35294-1170
This page took 0.02 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.