Author: Robert Hyatt
Date: 21:39:23 06/18/02
Go up one level in this thread
On June 18, 2002 at 19:02:45, Tom Kerrigan wrote: >So a common misconception seems to be that 64-bit chips would be twice as fast >as 32-bit chips. This is almost exactly like saying you can add small numbers >together twice as fast on an 11-digit calculator than on a 10-digit calculator. >(Computers spend most of their time working with < 32 bit numbers.) It doesn't >make any difference. The main advantage of 64-bit chips is that you can easily >address more memory and bigger files with them, and I've never heard of a 32-bit >chess program that was starved for memory. > >It makes sense that a bitboard chess program would benefit from the 64-bitiness >of a processor but I've never seen any evidence to support this. According to >SPEC, the Itanium doesn't run Crafty any faster than 32-bit programs, when >compared to a Pentium. (In fact, it runs Crafty a little slower.) Hyatt likes to >say that 64-bit chips are great for computer chess because the Alpha runs Crafty >really fast, but looking at SPEC, the Alpha runs everything fast. It only runs >Crafty a few percent faster than 32-bit programs. Two points: 1. The 21264 runs Crafty at > 800K nodes per second using a single processor at 600mhz. What 32 bit single processor can run crafty that fast? Answer: zero. 2. The Mckinley has run Crafty at 1.5M nodes per second at 1ghz. What 1ghz 32 bit processor can run crafty that fast? What 32 bit processor can run crafty at that speed _regardless_ of clock speed? Answer: nothing comes close. I have said it repeatedly... 32 bit programs won't profit a bit from a 64 bit architecture. They just end up pumping an extra 32 bits all over the cpu with no gain from doing so. 64 bit programs _will_ profit, because 32 bit machines require at _least_ two 32 bit operations to emulate a 64 bit operation. Two for ands/ors/xors, more for shifts, adds, etc. It doesn't particularly matter whether anyone believes this or not. Once the formal Mckinley numbers are published, the advantage will be quite obvious for 64 bit programs. Alphas are not as interesting any longer since they seem to be doomed. But until Mckinley, _nothing_ could touch an alpha (264) for Crafty, at least. I won't speculate on others. But I have _real_ data for the '264 and Crafty. > >A chip's design is MUCH more important than how wide its ALU is. Otherwise, a >386 would be more or less as fast as a Pentium 4 because they're both 32-bit >chips. Just because a chip is 64-bit doesn't mean it's fast. The Itanium is a >dog. So is the UltraSparc III. The ultrasparc is trash. All sparcs have been trash. If speed is the issue. And it has always been _the_ issue for me. Itaniums are fixing to take a huge jump upward in performance. A single 1ghz Mckinley is faster than my quad 700 xeon, even with the 1mb L2 cache for each processor on my machine. > >In other words, nobody should get fired up about 64-bit. Eh? _I_ am fired up about them. Because _I_ have run on them and see what they offer. > >BTW, 64-bit chips are very common. The Nintendo 64 had a MIPS R4000, which is >64-bit. Every RISC workstation for the past few years has been 64-bit, including >HP-PA, UltraSparc, POWER3/4, Alpha. > >The Itanium 1 is a dog. Everybody says the Itanium 2 will be fast, but according >to what Intel itself has disclosed about performance, it's not going to be >faster than a fast Pentium. Plus, it'll be really expensive. Personally, I think >the Itanium's design is stupid and am waiting for it to disappear. > >The AMD Hammer (successor to the Athlon) will rock. It's 64-bit, it has a deeper >pipeline, a better branch predictor, more registers (!), an on-die memory >controller, and more. Prerelease 800MHz Hammers are almost as fast as 1.6GHz >Athlons at Quake 3, and production Hammers should clock faster than Athlons. I >think the Hammer will be the best chip for computer chess for a long time to >come. > >-Tom Wouldn't argue a bit about itanium v1. It was horrible. V2 looks very impressive. Far more impressive than the AMD stuff to date. Whether the new AMD stuff will deliver performance like Mckinley will be speculation until one arrives. I generally get info about new chips pretty quickly since there are lots of chess nuts at the various chip makers. I will report on AMD numbers when something "printable" shows up... Mckinley numbers are not yet official, but I have seen the results personally and since I didn't have to sign a ND agreement, I have felt able to discuss them in general terms. 1.5M nodes per second for one cpu is quite fast... even if it is only clocked at a pedestrian 1ghz for now.
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.