Author: Keith Ian Price
Date: 17:26:38 04/20/05
Go up one level in this thread
On April 20, 2005 at 17:35:56, Frank E. Oldham wrote: >On April 20, 2005 at 14:41:49, Robert Hyatt wrote: > >>On April 20, 2005 at 12:33:24, Keith Ian Price wrote: >> >>> >>>IBM recently came out with their Power-5 module: >>> >>>http://static.userland.com/weblogsCom/images/wallyswisdomwarehouseweblogscom/8XPower5MCM.jpg >>> >>>This module has 4 dual-core multi-theaded Power PCs similar to the ones used in >>>the Mac G5. That makes for a total of 16 virtual cores, and IBM has a system >>>that ties 8 of these modules together with a 4GB/s bus for a total of 128 >>>virtual cores. The other four chips in the module are 4x36MB L3 cache. Since >>>Crafty already gets about 1500 kns on a fast processor, and the mult-threading >>>on a core offers about a 15-20% speedup, Crafty would likely exceed 100,000 kns >>>on a full system, especially if the hash tables could be kept in the L3 cache. >>>First, would it be possible to run a 128-thread version of Crafty? If so, do you >>>suppose that IBM might be interested in affording you the use of one of these, >>>as a Professor of Computer Science, to have a match against the self-proclaimed >>>successor to Deep Blue? I imagine they would get some good publicity having an >>>off-the-shelf IBM computer beat the specially designed chess computer in a >>>match. What do you think? >> >>Hard to say. The "dual-core" part sounds good. The other part about what >>appears to be a form of "hyper-threading" does not. HT for Crafty is actually a >>losing proposition after the changes Eugene and I worked on (with AMD) for the >>NUMA stuff last year. My dual xeon has HT disabled. >> >>But that aside, this could be a pretty powerful box. I've said all along that >>the FPGA approach is not a particularly attractive approach considering what >>could be done with an ASIC (ala' deep blue 2) vs a far slower FPGA solution. I >>would not be surprised if later this year the dual-core boxes were able to >>surpass the Hydra performance level, we will see... > >The SMT (hyperthreading) can be disabled -- they do this for some SPEC testing. >On a single-core basis,, the POWER5 at 1.9GHz is about 10% slower than an >Opteron at 2.5GHz on Spec Int type codes, >but about 100% faster on Spec fp codes. And IBM will sell you 64 cores... Runs >AIX or Linux 64-bit. > >My rough estimate is that crafty could enter the Deep Blue range of 100+ MNPS on >a large config. > >Frank That's what I thought. Now the only problem is to figure out how to get IBM to donate a big system to UAB for "research". It would take the "ft" right out of Crafty, and get it back to its roots... ;-)
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.