Author: Robert Hyatt
Date: 20:58:53 03/17/03
Go up one level in this thread
On March 17, 2003 at 16:57:28, Tom Kerrigan wrote: >On March 17, 2003 at 16:13:11, Robert Hyatt wrote: > >>Evidence that a machine with interleaved memory is faster than a machine without >>interleaved memory? I certainly had two such machines sitting side-by-side a >>few >>years ago, and it made a difference. >> >>I can't imagine a case where it would not, unless memory bandwidth isn't an >>issue for >>a program. For Crafty, at least, it certainly is. > >How do you know that memory bandwidth is important for Crafty? Others have >reported that Crafty scales linearly (1:1) with processor clock speed, which >indicates that it relies very little on main memory. You said you don't have the >appropriate machines to test Crafty's scaling.\ I said I don't _now_. I also said I _did_ have a while back. > >Also, what's so big and sequential in Crafty that memory bandwidth (and not >latency) is so critical? > I shouldn't have to explain this, but I will. Interleaving affects _both_. Because a PC doesn't just read 8 bytes and quit. Older machines had a 32 byte cache line. Newer machines stretch this to 64 and 128 depending on whether we talk about L1 or L2. Interleaving speeds up bandwidth, which also speeds up latency when the cache controller has to read _that_ much stuff... But I also have some high-bandwidth requirements for table probes for things like move generation, hash probes that read 48 bytes at one chunk, etc. Crafty doesn't scale "linearly" if the processor speeds up but memory does not. >-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.