Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Reply to Rajen Gupta re: 64 bit

Author: Vincent Diepeveen

Date: 14:09:01 07/25/03

Go up one level in this thread


On July 24, 2003 at 20:55:49, Tom Kerrigan wrote:

crafty GCC x86 32 bits mode GCC 3.3: 1094
crafty GCC x86-64 64 bits mode GCC 3.3: 1562

64 bits and more registers is faster: 43%
And we both know how bad GCC is for 16 registers when compared to other
compilers. The big wins still have to come...

see 32 bits:
http://www.specbench.org/osg/cpu2000/results/res2003q2/cpu2000-20030421-02102.html

see 64 bits:
http://www.specbench.org/osg/cpu2000/results/res2003q2/cpu2000-20030421-02093.html

>>>>On July 22, 2003 at 14:54:48, Rajen Gupta wrote:
>>>>
>>>>You are comparing x86 executables here.
>>>>That's like running 8 bits XT executables at the P4.
>>>>
>>>>Do you understand that?
>>>>
>>>>If not then it is useless to say that things will change for chess software a
>>>>lot when there is native good x86-64 compilers for opteron (and future intel
>>>>x86-64 platforms which does *not* include that 3.06Ghz Xeon MP).
>
>Not really.
>
>First of all, the XT had an 8088, which was a 16 bit processor.
>
>Second, XT executables use a lot of the x86's more CISC-ish instructions that
>run very poorly on modern processors, and they have data that isn't aligned
>right for modern processors, and they make heavy use of segmentation because of
>the tiny amount of flat-addressable memory, so of course they will run very
>poorly on modern processors.
>
>Using that as an analogy for x86 vs. AMD64 is a mistake because with the
>Opteron/Athlon 64, there is very little difference between how the processor
>works in each mode. Instructions are fetched, decoded, issued, executed, and
>retired in virtually the same way regardless of if it's 32 bit code or 64 bit
>code.
>
>There's a small performance penalty when running 64 bit code on an AMD64
>processor vs. 32 bit code because in 64 bit mode, pointers are all twice as big,
>which means a less efficient use of memory. Then again, there's a small
>performance gain from having the extra registers that are available in 64 bit
>mode. Overall, the performance gain from switching to 64 bit is reportedly
>10-25%, although recently an AMD64 build of SETI was released and it runs slower
>than the 32 bit version, by something like 10% if memory serves.
>
>So running regular x86 programs on an Opteron is in no way crippling it, because
>the instruction mix is already good, the data is aligned right, there's almost
>certainly no use of segmentation, and the only "bad" thing about it is that the
>extra 8 registers aren't available, which don't make THAT big a difference.
>
>Of course, chess is a somewhat special case, because some programs use bitboards
>and 64 bit math significantly speeds up bitboard operations. Of course, you
>wouldn't see nearly a factor of 2 speedup because bitboard operations are still
>a small percentage of instructions even for programs that make heavy use of
>them.
>
>-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.