Author: Robert Hyatt
Date: 18:27:30 09/03/01
Go up one level in this thread
On September 03, 2001 at 19:20:35, Vincent Diepeveen wrote: >On September 03, 2001 at 12:23:53, Robert Hyatt wrote: > >>On September 03, 2001 at 12:06:26, Vincent Diepeveen wrote: >> >>>On September 02, 2001 at 17:26:31, Tom Kerrigan wrote: >>> >>>>On September 01, 2001 at 08:40:46, Vincent Diepeveen wrote: >>>[snip] >>> >>>>>completely (especially price range), as well from the >>>>>servers from alpha and especially from SUN and all other cpu >>>>>brands when running 32 bits (computerchess) software, >>>> >>>>Seems like lots of computer chess programs are 64 bit now, and the Alpha can >>>>more than hold its own against any PC processor running either 32 or 64 bit >>>>code. >>> >>>This is what i mean, an alpha 2 processor is around $20000 in the >>>store now. that's for a 833Mhz dual alpha. >>> >>>A dual 1.2Ghz K7 of course blows that away for DIEP. >>> >>>An alpha 633Mhz (21164) used to be for diep around the same >>>speed like a PII at 380Mhz. >>> >>>And the PII was 17.3% faster than a PII >>>the K7 is another 10% faster than a P3 >>>the palomino i have now is another few % faster. >>> >>>21264 has about the same number of registers as a K7, >>>and the 21264 has huge costs for branch mispredictions. >>>The 4 instructions a clock it sometimes can do hardly make up >>>for the lack of 400Mhz in speed. >>> >>>And that for a factor of 10 in price difference... >>> >>>>-Tom >> >> >>Your numbers are simply wrong. You are using a poorly behaved program, >>written to work well in a 32 bit machine, to judge a 64 bit cpu. The 21264 >>for crafty will rip the cover off of any PIII/PIV you care to drag up, even >>though the alpha isn't running anywhere near the same clock rate. > >Of course i'm having a program which is not made for a 64 bits machine, >that's pretty logical. > >Note that crafty's numbers at the Sun aren't too great, though this IS >a 64 bits machine. DIEP at a sun is more than OK (considering its >Mhz). The sun is a piece of trash. It's performance is far behind the rest of the world, by a factor of 2-3 at least. And it isn't a true 64 bit machine either. In fact, so far, sun hasn't delivered a decent 64 bit compiler for the ultra-sparcs. And all the O/S code is _still_ 32 bit code. > >>Your program would also die on a Cray, because you haven't given any thought >>in how to use a Cray's unique features. But that doesn't mean _everybody_ is > >This is simply dead wrong, please tell me where new cray machines >are getting build. Isn't the company bankrupt or something? They are still shipping them daily. Visit cray's web site. Don't know where you get your info, but it is wrong. The state of alabama just bought a _new_ Cray in fact... > >AFAIK all crays build now use alpha processors, also a horse i wouldn't >bet on too long. This is wrong. > >So should i rewrite diep for a cray? No, but you should stop making pronouncements about various processors, based solely on how your program runs on them. That is _not_ a good benchmark. I'll bet you any amount of money you want that Cray Blitz, in pure fortran, will run 10x faster than your program, on a cray. Because that code was designed to work on a vector architecture, and has specific algorithms designed to take advantage of that kind of hardware. > >>in the same boat on 64 bit machines. And once you start to use the features >>they offer, the performance gains are impressive. >>The old saying "garbage in garbage out" still holds. You have to have a program >>that is designed for the architecture. > >Feel free to write crafty in VERILOG or whatever, and when you have >pressed a cpu, come back and then we'll play with the latest diep >version against your old crafty version which took so long to translate. > >Nah that's pretty mean from me to say. I don't have any idea what you are talking about. I'm not designing a chip here. I am writing code to exploit certain hardware features that all future cpus are going to have (64 bit words, for example, vector operations, most likely). > >To rewrite a program to a very specific machine which focuses upon >bandwidth, because let's not forget that Crays are made for bandwidth, >NOT with computerchess in mind, it takes years to translate a couple of >megabytes of C code to Cray assembly! You don't have to do it in assembly. You just have to understand the architecture. Which is _exactly_ what you are doing for the PC. Except you are focusing specifically on the X86, which is doomed to be replaced by 64 bit processors eventually. > >That's not only a waste of time, but also a very dumb thing to do. > >At the time you have rewritten it, your program is outdated! >OUTDATED! Not if done right. I can maintain assembly code about as well as I can maintain C code. Did it for years in my operating systems work. Never had a moment's problem with millions of lines of assembly.
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.