Author: Vincent Diepeveen
Date: 06:59:21 01/03/03
Go up one level in this thread
On January 02, 2003 at 21:25:31, Robert Hyatt wrote: current crays are 1ghz clocked and have 256K cache too. 16 x 1Ghz = 16Ghz If you get 6-7MLN at that, then i do not find that very impressive when compared to crafty which will get easily a similar node count and it is as you say not optimized for Cray. >On January 01, 2003 at 11:58:57, Vincent Diepeveen wrote: > >>On January 01, 2003 at 11:53:46, Robert Hyatt wrote: >> >>>On January 01, 2003 at 02:01:15, Vincent Diepeveen wrote: >>> >>>>On December 31, 2002 at 10:58:40, Robert Hyatt wrote: >>>> >>>>>On December 31, 2002 at 08:47:37, Frank Phillips wrote: >>>>> >>>>>>On December 30, 2002 at 19:25:09, Vincent Diepeveen wrote: >>>>>> >>>>>>>On December 30, 2002 at 13:34:31, Frank Phillips wrote: >>>>>>> >>>>>>>>On December 30, 2002 at 11:33:18, Vincent Diepeveen wrote: >>>>>>>> >>>>>>>>>On December 30, 2002 at 05:26:32, Graham Laight wrote: >>>>>>>>> >>>>>>>>>>See http://www.talkchess.com/forums/2/message.html?54285 in the other forum. >>>>>>>>>> >>>>>>>>>>-g >>>>>>>>> >>>>>>>>>it is vector CPU's. Not comparable with cpu's that do things like computerchess >>>>>>>>>at all. So for computerchess that machine isn't that fast at all. >>>>>>>> >>>>>>>>Wasn't the Cray a vector machine? Running Cray Blitz by Hyatt et al. >>>>>>> >>>>>>>Yes. 16 processors in total got him to about 500k nodes a second. >>>>>>> >>>>>>>I do not know what Mhz Cray Blitz ran on. But probably Hyatt can enlighten >>>>>>>us about it. >>>>>>> >>>>>>>However for matrix calculations and such that Cray was >>>>>>>considerably faster than it was for Cray Blitz. >>>>>>> >>>>>>>Then you'll see the Cray didn't do that impressive for each >>>>>>>Mhz whereas it was a lot more impressive for vector processing. >>>>>>> >>>>>>>Compare both Mhz of todays x86 with the Cray times 16 back then >>>>>>>and the vector power versus todays x86 and you'll know what we are >>>>>>>speaking about. >>>>>>> >>>>>>>Best regards, >>>>>>>Vincent >>>>>> >>>>>> >>>>>>No I cannot. I can see that it might be slower MHz for MHz, but given its >>>>>>awesome speed (35 trillion calculations per second) I would have thought it >>>>>>would be a very strong chess machine, particularly if the program was written >>>>>>with vector processing in mind. >>>>>> >>>>>>Frank >>>>> >>>>> >>>>>Of course it would. But you have to: >>>>> >>>>>(1) be willing to expend the effort; >>>>> >>>>>(2) understand vector processing or else put forth the effort to figure out >>>>>how it might apply to chess; >>>>> >>>>>(3) not write everything off as "impossible" just because you don't know how >>>>>to do it _now_. >>>>> >>>>>(4) be willing to spend a lot of time "getting into vector processing mode" >>>>>and learn how to use it effectively. It is just like "getting into bitmaps". >>>>>_some_ are simply incapable of doing so... >>>> >>>>You didn't do all that for Crafty. Otherwise even the current 1Ghz McKinley >>>>would be 50% faster than Alpha and you just posted it isn't. >>>> >>>>How comes? >>>> >>>>Happy programming in 2003, >>>>Vincent >>> >>> >>>Mckinly ain't a vector machine. Not even close. >>> >>>So, once again, I have no idea what you are talking about. (Is this >>>becoming a _common_ comment by me and others?) >>> >>>BTW, I didn't say I vectorized Crafty. I _did_ say I vectorized Cray Blitz >>>and it ran like blazes on a vector machine. Itanium is not a vector machine. >>>No X86 lookalike is a vector machine either. >>> >>>If you don't know what a vector machine is, find a good book. Or I can give >>>some sample code for the Cray (assembly) to show what it is about. >> >>Cray blitz was like 500k nps at 16 processors. > >On a C90 with 16 processors, correct. We ran on this machine in 1992, when >it was first available to us. > >> >>16 * 100Mhz Cray VECTOR processor (capable of 29 integer instructions >>a clock or so versus x86 about 3). = 1.6Ghz > >The C90 was not a 100mhz processor. The clock cycle time was 4 >nanoseconds, which is 250mhz if you want to use a wrong measure. > >And the "capable of 29 integer instructions a clock" is total nonsense as >usual. Each CPU could execute _one_ integer instruction per clock. _never_ >any more than that. In vector mode each cpu could do four reads and two writes >per cycle along with as many vector operations as you could chain together, >so that it _could_ do 32 (or more) adds/subtracts/multiples/etc per cpu, per >clock cycle... > >Of course, if you would just look at the cray architecture, you would know this >without resorting to wrong statements... > > >However, the T90 was twice as fast, with twice as many processors, and it >brought a few new things to the table including (for the first time) a real >cache for scalar variables, something no prior Cray ever had. And as I >reported here at least a year or so ago, I played Crafty on my Quad 700 vs >the T90 and the T90 was hitting around 6-7M nodes per second, typically. > >6-7M nodes per second in 1995 was _not_ bad (that was when the T90 came >along, although we had no ACM events after 1994 so it never played in a >computer chess tournament. > > > > > >> >>You get more nps than that with crafty at a 1.6Ghz K7. > >So? Do you understand the concept of time? C90 = 1992. 1.6ghz K7 = 2002. >ten years (or more) apart. Go back to 1992 and find _any_ computer that could >search 500K nodes per second, excepting Deep Thought. Go back to 1995 and >find any machine that could hit 6-7M except for Deep Blue. > > >> >>You're doing around 1 MLN a second with crafty at a 1.6ghz K7 >>at the great vector processor which a Cray is you did 500k NPS. >> >>Best regards, >>Vincent > > > >So? I did a lot of things in Cray Blitz that I can't afford to do in Crafty >just yet. That 500K nodes per second was 1992 hardware. This is 2002. Why >compare today's PC to a supercomputer 10 years old? > >Oh yes, you really have no idea what a supercomputer is, so time has little >meaning _either_... > >Just a room full of monkeys typing on keyboards that occasionally put something >together that sounds like it might be legitimate, when it is far off the real >mark..
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.