Author: Robert Hyatt
Date: 06:31:06 03/25/02
Go up one level in this thread
On March 25, 2002 at 07:05:26, Vincent Diepeveen wrote: >On March 24, 2002 at 21:42:32, Robert Hyatt wrote: > >>On March 24, 2002 at 21:10:57, Eugene Nalimov wrote: >> >>> >>>We are talking about *absolute* performance, not about "nodes per MHz", right? >>>Right now I can buy 2.2GHz x86 system, but I believe that fastest Alpha is 1GHz. >>> >>>I just run Crafty's "bench" on my 1.13Hz notebook and got 620knps, and I think >>>that fastest x86 system will be faster than fastest Alpha. >>> >>>Eugene >> >>First, several mis-matches. >> >>1. I used gcc to compile Crafty, which Tim used to compile on the alpha. >>So there is no compiler differences that are very significant... > >>2. I believe that Vincent has been saying that 64 bit processors simply >>don't help. This is incorrect. They help significantly. There is no >>technical reason why a 64 bit processor can't be clocked just as fast as >>a 32 bit processor. The marketing force is simply not driving the development > >In a year or 10 you may be right, for now not a single 64 bits >processor is clocked *near* 32 bits processors. > >You are still busy with stupid old 600Mhz 21264 processors. > >I tested myself on the mothers. They suck ass for computerchess, the >only interesting is absolute performance, not how fast they are for >600Mhz, because in CPU design the L1 and L2 cache are very important. >If they didn't manage to clock those at 1.6Ghz, then that's their problem, >not mine. The 600mhz alpha produced 800K nodes per second for me. What processor do you propose to do the same in the intel world. Knowing that 1ghz alphas are available? That hardly "sucks". Unless you mean "it sucks the doors off those 32 bit processors..." > >For FPU, that's another story, that's not computer chess. Crafty doesn't use the FPU. It is a pure integer program... > >>that way _yet_. But if someone wants to make a fast 64 bit chip, and believes >>that they can market it to make a profit, then it will happen. > >Oh they will come, no doubts here that it will happen. Look at the >McKinley design. Of course i can't pay a mckinley, but you could put >perhaps 16 in parallel on a supercomputer. Will be perhaps even >cheaper than the 10 million dollars for a 16 processor 21264 system. you are really off on the price. By more than a factor of 10... > >Of course clocked at 1Ghz each. > >>3. If you use gcc, your 1.13ghz notebook will be well off that pace, and >>that will put it back into the light of a relatively equal comparison since >>we would all be using the same compiler. > >This is dead wrong of course. you use the compiler that's fastest for >the processor. If you look well at the specbench.org site you'll see >that they also quote which compiler is used. What I said was not "dead wrong". DEC has a compiler for the alpha that is much better. It is as much better than GCC as MSVC is better than gcc on the intel platforms. I simply chose to use the _same_ compiler so that any difference was due to the hardware, and not the optimizer... > >Compaq C V6.4-214-46B59 >Program Analysis Tools V2.0 >Spike V5.2 DTK (1.461 46B5P) >Compaq C++ V6.3-010-46B2F > >So they first analyzed the program and only after that recompiled it >in order to get faster. > >It is very EASY to see that this is the fastest compiler for the Alpha. > >Amazingly for the XP2000 used was the intel compiler. Even though i >don't trust this compiler at all for my software. > >>I was personally far more interested in the nps per megahertz since that seemed >>to be the point vincent was attacking the alpha on, and the numbers don't >>support his argument at all... > >No that's not interesting Bob. You know it and i know that they can >easily clock L1/L2 cache at 1Ghz nowadays, but try to clock it at >2Ghz!! L1/L2 at 2ghz is not a problem. SRAM can run as fast as you want it. DRAM is another thing however.. > >Of course a machine at 2.2Ghz or 1.67Ghz is going to have more problems >clocking the caches higher. > >The only INTERESTING thing is how fast a processor + compiler performs >for a program. If you build a 1Ghz processor, then it gotta beat >a 2Ghz K7 simply. If it doesn't, THEN YOU ARE SLOWER. For Crafty, it _does_ beat the K7 when running on a 1ghz alpha... that was my point. > >>>>I don't know the specifics about the dual, but it was only 1.5X faster. We >>>>later improved this a lot as the "lock" facility we used to start with was >>>>very slow on the alpha architecture. >>>> >>>> >>>> >>>> >>>>> >>>>>Relative to the 1.67Ghz from the K7 the alpha achieves like >>>>>a 102/122 x 1.667Ghz = 1.394Ghz K7 >>>> >>>> >>>>Show me that 600mhz K7 that can do .8M nodes per second with crafty... > >Show me the alpha that can beat a 1.67Ghz XP, YOU WILL NOT FIND IT. >In fact it is 30% slower even for CRAFTY which is supposed to profit >from 64 bits. OK... first let's stick with the 600mhz 21264 before going to the 1ghz version, since I have actual data from that processor. Do you think that XP will do 800K? I don't. Dual 1.4 athlons did about 1M nodes per sec last time I tried. Dual 600 alphas did 1.2M. The 1ghz parts are faster but I have not actually run on them personally, and only have a small bit of data from others that have. > >So a 32 bits processor doing 3 instructions a clock >is kicking the HELL out of a 64 bits processor doing 4 instructions a clock, >the 32 bits processor being clocked higher only 60%. You've yet to show me _any_ 32 bit processor that is "kicking the hell" out of an alpha... > >VERY HARD DATA!! > >Alpha sucks simply bigtime for the kind of integer operations a chess >program is using, that's a *logical* conclusion to draw. Not for the kind of integer operations _I_ do... > >And yes, the alpha compiler is using many instructions to remove branches, >the X86 guys didn't invent those themselves... > > > >>>>Then I'll be a believer, not now... >>>> >>>> >>>>> >>>>>In theory the 4 instructions a clock for alpha versus >>>>>3 instructions a clock for K7 give 33% speedup: >>>>> >>>>>1.000 Ghz + 33% = 1.333Ghz >>>>> >>>>>It achieves however 1.394Ghz >>>>> >>>>>In short i am missing the speedup for being 64 bits at all! >>>> >>>> >>>>Because you are looking at the wrong data.. :) >>>> >>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>>> >>>>>> >>>>>>>> >>>>>>>>>-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.