Author: Keith Evans
Date: 23:45:56 02/20/04
Go up one level in this thread
On February 21, 2004 at 01:19:46, Robert Hyatt wrote: >On February 20, 2004 at 21:15:10, Bob Durrett wrote: > >>On February 19, 2004 at 13:18:03, Robert Hyatt wrote: >> >>>On February 18, 2004 at 16:43:23, Jorge Pichard wrote: >>> >>>>On February 18, 2004 at 16:00:38, Robert Hyatt wrote: >>>> >>>>>On February 18, 2004 at 15:43:17, Aaron Gordon wrote: >>>>> >>>>>>On February 18, 2004 at 15:34:32, Mark Young wrote: >>>>>> >>>>>>>Hydra gets suckerd by some nice anti-computer play! A very ugly game by Hydra. >>>>>> >>>>>>Doubt Hyatt would want to work with the Hydra programmer. I believe it was >>>>>>Donninger (lead programmer?) who said, "The only good American is a dead one" or >>>>>>something to that effect. >>>>> >>>>> >>>>>Correct. But then, I'm not sure why the subject "Hydra needs hyatt on team" >>>>>came up either. I'd suspect they can do just fine without me. I'm not really >>>>>interested in looking at hardware solutions when we have such good >>>>>general-purpose "solutions" like the opteron around. :) >>>> >>>> >>>>I was just thinking about the comparison of two of those Hydra cards with a Dual >>>>Opteron 248 or 4 Cards versus Quad Opteron. I believe that Crafty using the same >>>>Quad Opteron that you previously use, is much faster than Hydra with 4 FPGA >>>>cards :-) >>>> >>>>Jorge >>> >>> >>>It is hard to say. I know how fast I was running. But their parallel >>>implementation has some issues to deal with that I get to ignore, and some of >>>those issues are pretty important, such as accessing the hash table, >>>communicating with other processors, etc. Fast on a SMP box, not so fast on a >>>box that depends on the PCI bus to talk to the FPGA cards... >> >>Is there no alternative to a PCI bus? If not, maybe we need to invent >>something. >> >>Bob D. > > >Not on a PC. That is the only interface there is for the moment. Other >machines have something better, but they cost a _lot_ more... A couple of people have tried DIMMs, but it's problematic. Intel bought a company Nuron that was doing it to accelerate public key crypto, and there's a Hong Kong university that's tried - see Pilchard. Don't know anything about it, but see http://www.tik.ee.ethz.ch/tik/education/lectures/RC/WS03_04/Plessl_TKDM.pdf too. One huge problem is many motherboards have very few slots, and even worse people tend to scramble the data lines to make routing easier. And the DIMM boards have to be relatively small, and not too power hungry. -K
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.