Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: News about Alpha 21464 (EV8)

Author: Bo Persson

Date: 13:32:24 01/22/01

Go up one level in this thread


On January 22, 2001 at 16:12:41, Vincent Diepeveen wrote:

>On January 22, 2001 at 09:52:01, Robert Hyatt wrote:
>
>>On January 22, 2001 at 03:10:18, Gregor Overney wrote:
>>
[...]
>>
>>
>>why would you conclude that more instructions per cycle slows the cycle
>>down?  Because the Intel does it?  Different animals.  Pentium = 64 bit
>>bus.  Alpha=256.  With much higher memory bandwidth to boot..

Exactly, the Alpha instruction set was also designed for this.

Intel PIII/PII/Pentium got theirs from the 486/386, which was an
extension of the 286, which was an extension of the 8086, which...

And then you add MMX, SSE, etc. A total mess!

Surprising that Intel can get *any* performance out of this.  :-)

>The more instructions a clock you do the bigger the problem is
>when you have a lot of branches in the program as currently processors
>need to reload all the instructions.
>
>I guess gregor is referring to
>that where on the PIII/PII it can execute 3 instructions a clock
>you have a 10 penalty for each misprediction
>
>With P4 which can issue 4 on paper (in reality it has problems
>which should prevent it from executing 4 simple integer instructions
>a clock) but already has a branch misprediction penalty of 20 clocks
>which is hell of a lot.

But...

The 20 stage pipeline is run on a double clocked ALU, so 20 internal
clocks is again 10 external clocks!

It also has a trace cache for decoded instructions, so the pipeline
doesn't have to do *all* the work again.


Bo Persson
bop@malmo.mail.telia.com



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.