Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Latency versus Information Bandwidth: Questions

Author: Matt Taylor

Date: 11:41:56 12/05/02

Go up one level in this thread


On December 05, 2002 at 10:25:48, Robert Hyatt wrote:

>On December 05, 2002 at 01:16:46, Jeremiah Penery wrote:
>
>>On December 04, 2002 at 23:15:03, Robert Hyatt wrote:
>>
>>>On December 04, 2002 at 20:29:52, Bob Durrett wrote:
>>>
>>>>
>>>>The recent threads shed some light on the issue of when one is more important
>>>>than another, but the answer is sketchy and seems to be "depends."
>>>>
>>>>For current chess-playing programs, which is more important?  Latency or
>>>>bandwidth?  Why?
>>>>
>>>>Is the answer different if multiple processors are used?
>>>>
>>>>Bob D.
>>>
>>>First, chess engines have a bad habit of doing real random-access probes to
>>>memory for things like hashing.  That is latency-dependent rather than bandwidth
>>>because we are not reading large chunks, but small scattered 16-32 byte
>>>blocks...
>>>
>>>That said, programs do need some bandwidth as you have to keep the cpu fed with
>>>instructions and data and that stuff resides in memory unless the program is
>>>small enough to tuck away in L1/L2/L3 cache.
>>>
>>>multiple processors increases the bandwidth requirement.  Two cpus require
>>>twice the bandwidth as one, and generally, high-end server boards provide two-
>>>way memory interleaving to double the bandwidth.
>>>
>>>Given the choice of 2x the bandwidth or 1/2 the latency, I'd go for reducing
>>>the latency.  But that is basically impossible.  Latency as been stuck at
>>>100-120ns for 20+ years now...
>>
>>I'm not sure why you persist in giving this 100-120ns number, when several
>>credible sources have said that the figure of today is no more than 75ns.
>
>
>Did you see the thread on rec.games.chess.computer?  75ns is often quoted but
>that is the
>latency in the SIMM/DIMM.  That is not the _complete_ path from the CPU to RAM.
>
>100 is *very* good.  120-135 is much more typical.
>
>If you can find a box that shows 75ms latency for random reads of one byte, I'd
>certainly like
>to see it.  Don't do random reads of 128 bytes as that amortizes latency over
>several words
>of data and makes it look better...  I want to know exactly how long I have to
>wait after I
>do something like mov eax, [ebx] before I can use that result in eax.

I think most DDR-based systems have about 100ns latency. Opteron is quoted as
80ns latency -total-. Just another reason to look forward...

http://www.amd.com/us-en/assets/content_type/DownloadableAssets/Optimization_-_Tim_Wilkens.pdf

-Matt



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.