Author: Christophe Theron
Date: 09:52:42 04/04/00
Go up one level in this thread
On April 04, 2000 at 04:14:42, Tom Kerrigan wrote:
>On April 03, 2000 at 22:48:46, Christophe Theron wrote:
>
>>On April 03, 2000 at 22:08:49, Tom Kerrigan wrote:
>>
>>>On April 03, 2000 at 20:10:00, Christophe Theron wrote:
>>>
>>>>I would suggest to go for the Athlon. The PIII really sucks, for my program at
>>>>least. I don't know why, but the PIII-450 is slower than the K6-2 400 for Chess
>>>>Tiger (close to 10% slower!).
>>>
>>>The K6-2 has much shorter pipelines than the PIII. That makes it significantly
>>>faster, but also keeps the clock speeds relatively low.
>>
>>
>>I suppose this explains why the K6-2 can execute more instruction per cycle than
>>the PIII: each pipeline "step" in the K6-2 does more work. On the other hand,
>>all this work to do means that the clock speed cannot be pushed too far. Is this
>>what you mean?
>
>Exactly. I think the main problem is with branch mispredictions. If you start
>speculatively executing instructions after a branch, and you guessed wrong, you
>have to flush the pipes and start over. Restarting a 12-stage pipeline really
>kills you because you have to wait 12 cycles for the first instruction to
>finish. I've heard that the PIII wastes 30% of its time due to mispredictions.
>
>>>The reason the Athlon may be better than the PIII for computer chess is the huge
>>>L1 caches. I suspect the integer core of the Athlon is pretty much equivalent to
>>>the PIII's.
>>
>>
>>Athlon or K6-2, it is the same thing for Tiger (at the same clock speed I mean).
>>So I'm not sure the L1 cache makes a difference here. And, at least for my
>>program, it seems that the Athlon integer core is better than the PIII's.
>
>I can't remember how much L1 cache the K6-2 has, but the Athlon has much more
>(128k). The Athlon also has a much faster L2 cache. So the K6-2 might be
>inherently faster than an Athlon, but once in a while it gets stuck waiting on
>memory, so they end up the same.
>
>-Tom
As far as I know the K6-2 has 32K for code and 32K for data.
So probably most of the core of my engine fits into this already, and having
more L1 cache doesn't help...
Christophe
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.