Author: Vincent Diepeveen
Date: 11:54:13 10/02/02
Go up one level in this thread
On October 02, 2002 at 14:23:53, Tom Kerrigan wrote: >On October 02, 2002 at 11:56:11, Vincent Diepeveen wrote: > >>Most supercomputer processors are stupid alpha's or stupid other cpu's >>which run >> a) at lower Mhz speeds >> b) give lower performance > >Wait a minute, something just occurred to me. What clock speed does the Alpha >run at? 1.25GHz. What clock speed does McKinley run at? Last I heard, 1GHz, at >most. So not only are you wrong on the performance count, you're also wrong on >the clock speed count. >-Tom I do not know a single supercomputer with alpha 21264 at 1.25 Ghz. Correct me if i'm wrong. If i browse around at top500.org i see: 3 alpha 1 Ghz 4 alpha 1 Ghz 6 alpha 1 Ghz 37 alpha 1 Ghz 38 alpha 1 Ghz I'm not seeing a single one clocked higher than 1 Ghz. Also the alpha is dead. the McKinley takes over! For me Mckinley is way faster than alpha. Alpha 21264c for me is slower than K7. Not much. But still. McKinley is 33% faster. That's without optimizing for 64 bits yet, where both alpha + mckinley will profit. though it won't bring much :) Look at specs: alpha : 4 instructions a clock mckinley : 6 instructions a clock So *potential* the mckinley is 50% faster. Of course that never will get achieved, we must be realistic. Doing more than 6 instructions a clock is really nonsense for chess. We still didn't talk about things like branch mispredictions. The alpha has very big penalties for a branch misprediction. i don't know about mckinley but seeing the results i had at the 2 cpu's i bet the mckinley is either having less or handling it better somehow. So i don't know what you care for, but i care for what i have SEEN. I see that the mckinley is with an edition 1.0 compiler already way faster than any other cpu i ever have seen. Now i don't know how hard guys like nalimov work, but obviously the intel c++ team that made that cross compiler which i used to benchmark diep on mckinley, they have a big disadvantage in time compared to the alpha team who already could work on the alpha compiler for a quarter of a century or so (?). Very obviously that is very promising for newer editions of their thing. The achieved results i have for the mckinley (so the 33% faster results) i did about half a year ago at a mckinley with a intel c++ 5.01 cross compiler which compiles for the itanium 1, not the itanium 2. I can list another set of disadvantages for the mckinley. I don't care. the important thing is that it's faster. the measurement is 33% faster. it was a very *accurate* measurement. Within the 1% limit even. The machine was completely idling when doing this test. It tells me that the thing could get way faster for me by optimizing and using newer compiler editions. When was intel c++ 5.01 released, long time ago i bet? Whatever statement on my behalf here, it's very obvious that the mckinley is going to have a superb future.
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.