Author: Robert Hyatt
Date: 20:14:32 11/15/97
Go up one level in this thread
On November 15, 1997 at 22:31:15, Keith Ian Price wrote: >On November 15, 1997 at 14:39:31, Robert Hyatt wrote: > >>On November 15, 1997 at 04:40:15, Keith Ian Price wrote: >> >>>On November 14, 1997 at 22:23:20, Robert Hyatt wrote: >>> >>>>On November 14, 1997 at 21:22:36, Keith Ian Price wrote: >>>> >>>>>On November 14, 1997 at 08:18:06, Robert Hyatt wrote: >>>>> >>>>>>On November 14, 1997 at 01:52:31, Keith Ian Price wrote: >>>>>> >>>>>>>On November 13, 1997 at 12:19:31, Chris Carson wrote: >>>>>>> >>>>>>>>see article: http://www.currents.net/newstoday/97/11/11/news5.html >>>>>>>> >>>>>>>>Interesting article. Lots of power 400 to 700 MHZ, >>>>>>>>Lots of RAM (up to 3 gigabytes). >>>>>>>>128 bit cache (up to 4 megabytes L3) >>>>>>>>64 bit power >>>>>>>>Native Windows NT 5.0, no translation needed. >>>>>>>> >>>>>>>>I hope this puts pressure on Intel!! :) >>>>>>>> >>>>>>>>Looks like 1998 might bring some competition. >>>>>>>> >>>>>>>>Best Regards, >>>>>>>>Chris Carson >>>>>>> >>>>>>>Intel plans to have a 450 Mhz Pentium II by the end of 1998 with a >>>>>>>100Mhz bus and 2MB of 450 MHz L2 cache. With this size and speed of >>>>>>>cache, I think that for computer chess, this Intel machine will easily >>>>>>>outdistance a 600Mhz cheap Alpha, and would be price competitive with >>>>>>>the 600 Mhz Alphas from DEC at that time. Which should be half the price >>>>>>>they are today. Merced--IA64--due in 1999, should reach 1 Ghz by the end >>>>>>>of 2000. This is, of course, still an unknown player. I think they feel >>>>>>>the pressure. >>>>>>> >>>>>>>kp >>>>>> >>>>>> >>>>>>Of course, within the year DEC's 21264 will be out, so comparing with >>>>>>the >>>>>>PII/450 is a poor comparison. How would a PII/450 compare to a 21264 at >>>>>>one gigahertz? >>>>> >>>>>I thought we were talking about cheap Alphas. Not $10,000+ Alphas. With >>>>>price and compatibility taken into consideration, I'd take the PII 450, >>>>>which will be more than half the speed, most likely. Almost doubling >>>>>speed for an additional $5000-8000 is not an option for me. Which is why >>>>>600 Mhz Alphas don't really compare to a $2500 PII-300 today. >>>>> >>>>>kp >>>>> >>>> >>>>Notice that you can go buy a 533 mhz alpha for under 5,000 bucks right >>>>now, >>>>which is about what you'd pay for a good PII/300. >>> >>>I can get a very nice PII-300 for $2499 from Quantex right now, with 8.4 >>>gig 9ms hard drive and 64 MB. The new LX chipset and AGP card are also >>>included with 17-in.monitor. And weren't you talking about the 21264? >>> >> >>Depends. If you are talking about a PII/300, then I'm talking about a >>21164, which I can buy for not quite 2x what you paid, and which is more >>than twice as fast. > >But only on Crafty. And 2X speed for 2X money is a bad deal on >higher-end computers unless you're Dollar Bill Gates. 2X speed doesn't >even get you another ply, so performance at today's levels requires 10X >speed for it to be worth 2X $2500. again, this depends. In typical middlegame positions, 2x will get another ply well over 50% of the time. The typical branching factor is 2.5-3x for most programs today, due to forward pruning algorithms... However, 2x performance for 2x the price is an *outstanding* deal. For reference Call Cray Research and ask about the difference in pricing between a 2 nanosecond machine (500mhz) and a 1 nanosecond machine(1000mhz). It generally costs a factor of at *least* 10x to go 2x faster, on the top-end of computer product lines. > >>But you also mentioned a PII/450 next year, and for that I'd compare it >>to >>the 21264/1000, or maybe even only /800, which will likely be available >>at >>less than 2x the cost atain, and provide more like 3x the speed. >> >>Poing being Intel isn't the only company. I love the P6. I'll love the >>P6 >>when it is re-released to replace th PII, early next year. But they >>don't >>come anywhere near the alpha. And when you factor price/performance, >>the >>alpha still looks real good. IE 2x the PII/300 performance at less than >>2X the cost. > > >But 2X is not enough for chess. And 3X is not worth an extra $6000 for >the 21264. 2x is not bad. In endgames it can be 2-3 plies. In the middlegame it is a ply at least 50% of the time. That is a measurable improvement... > >>>>And running Crafty, that >>>>533 mhz alpha would be about twice as fast. >>> >>>The 533 Mhz Alpha would definitely not be twice as fast as the PII-300 >>>running anything but Crafty, and I was pretty sure I heard you tell >>>Chris Whittington that it was only 20-25% faster than his PII-300, not >>>twice as fast. >> >> >>Sorry, but we got better data. Here are the *real* numbers, just for >>discussion: our "benchmark" suite of 6 positions, 2 each from opening, >>middlegame and endgame, averages 80K nodes per second on our P6/200's. >>On the 500mhz alpha, this same test suite runs at 250K nodes per second >>average. which is 3.1X faster. >> >>The PII/300 we have here (running both linux and NT) is not quite 1.5X >>faster than my P6/200, which is about what is expected. But this >>*still* >>leaves the alpha at 2x faster. >> >>The first numbers we discussed were numbers Bruce provided (/533 = 1.75X >>faster than his P6/200) but were based on a program that does almost >>everything >>in 32bit mode, while crafty does almost everything with 64bit words. I >>get a >>lot more. > >So you agree, only Crafty would improve by the 2X speed. I can't conclude this. I can only say that crafty certainly improves by that much. Others may move toward 64bit algorithms when 64bit hardware is more available. A 32 bit program isn't going to get any advantage from pumping 64 bits, which makes the alpha unattractive. But this will change, because chess programmers aren't stupid, and they will recognize, just as I did, that 64 bit algorithms are the way of the future. Remember when people were complaining about the 16 bit programs and how they were a dog on the P6? And most programs are now 32bit? So wouldn't you expect these same programmers to put 2 and 2 together and get 64? :) > >>> >>>>IE the PII/300 is not quite 1.5X faster than the P6/200 due to the poor >>>>L2 cache clock speed. The alpha/500 is 3.1 times faster than the P6/200 >>>>running Crafty. >>> >>>Is this what you told Chris? I thought Bruce said that it was 1.5-1.75 >>>times as fast as the P6-200, running Ferret. >> >>this is correct. I didn't have the /500 numbers at the time. Jason ran >>these tests after getting to Paris and playing around with compiler >>options >>on our 500 mhz alpha over there. the 533 should be 5% faster than this >>or >>so, based on clock speed... >> >> >>> >>>>By the time you can buy a PII/450, you will certainly be able to buy an >>>>alpha/600 for the same price-level, and run faster. >>> >>>I think the PII-450 will be faster than the Alpha 600 running most chess >>>programs, probably even Crafty. The 2MB of L2 cache at 450 Mhz is more >>>important than the extra 150 Mhz. If the Alpha is slightly faster >>>(running the one chess program I will be able to compile >>>natively-Crafty), it won't be noticeable, but the software >>>non-compatibility will be. >> >>I disagree, but we'll have to wait to see. But the PII/300 is less than >>1/2 >>as fast as the 500mhz alpha. So it has a ways to catch up... > > > >>>>And it won't be >>>>long, >>>>after the 1ghz alphas come out, that the 750mhz version will be >>>>price-competitive with Intel. >>> >>>And when Merced comes out, the PII-300 will be entry level for $1000 >>>bucks, so what? >>> >>> >>>>And then there are the PPC's which are already faster than the pentium >>>>machines, unfortunately, and getting faster. They will be at 1ghz >>>>before >>>>the alpha apparently... >>> >>>The G3 has a better shot than the Alpha. But it's still a long shot. I'm >>>not an Intel supporter. I've been programming computers since 1968, and >>>the first Intel machine I bought was a 486. I've always been a >>>value-for-money person, and right now, Intel or K6 is it. I'd love it if >>>Digital could come out with a machine that blows the doors off anything >>>else. I used to love the old PDP-11 I had, and I used to run Checkmo, a >>>DECUS program that played horrible chess on a PDP-8, but it was fun to >>>watch. I'll buy whatever I think represents the best value. >>> >>>kp >> >>They have the machine you describe, because nothing Intel has is even >>close to >>the alpha. But it is pricier. Of course part of this is caused by the >>fact >>that the intel products (at present) pump 64 bits over the bus to feed >>32bit >>superscalar execution units. The alpha pumps 128 and 256 bits over the >>buss >>to feed 64bit superscalar units. Merced is not going to be a linear >>increase >>in cost. Doubling the buss width to feed a real 64bit processor is not >>going >>to come cheap, and will narrow the gap between intel and digital pricing >>I >>suspect. How much it closes the performance gap (if any) is unknown at >>present of course... > >Still, the PII-450 will be at least 10X times my current P150s, which >makes it worth the extra cost over the cheap 150 non-MMX units I just >bought in September. 2 or 3X isn't worth thousands of dollars to me. >Just a difference in perception, I guess. Then again, I don't have to >try to win the WMCCC next year, so I can afford to wait. Nice discussing >things with you, Bob. > >kp it's an interesting question. when you say 10x I assume you mean non-mmx 150's vs a PII/450? Which sounds pretty reasonable, if not a little optimistic. My only cost comment is that from my perspective, the PII/300 is less than 1/2 the speed of the alpha/500, which are not 2x more expensive. The difference between 2500 and 4000 is not that great. the /500 machines will continue to drop, and they still have that 2x speed advantage. It is at least worth thinking about. IE Bruce bought one for Ferret... And it screams...
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.