Author: Vincent Diepeveen
Date: 10:25:36 10/25/02
Go up one level in this thread
On October 24, 2002 at 14:01:52, Anthony Cozzie wrote: >On October 24, 2002 at 05:11:19, Jorge Pichard wrote: > >>If you compare the architecture advantages of a Dual AMD MP such as: The >>operation per clock cycle, the floating point pipelines and the L1 Cache size >>you can immediately see why AMD is the King. >> >>http://www.amd.com/gb-uk/Processors/TechnicalResources/0,,30_182_865_4362,00.html > >1. Intel's x87 FPU is (shudder) stack based. Go look at spec fp 2k scores and >watch a 1GHZ sun blow away a 3GHz PIV. Intel made a deliberate decision to >optimize for multimedia (read: DSP) for the PIV; the PIV emphasizes SSE as a way >around the asinine stack based floating point unit. But a K7 or McKinley blows away that Sun. >2. PIV has a LARGER L1 Icache than athlon. What idiot tacked an 8K DIRECT >MAPPED L1 Dcache on it I have no idea. Also note it is 64 bits based that cache. So only 1024 words. The reason is very obvious of course. With a bigger L1 dcache they would not be able to clock the cpu to 3Ghz. Instead they would be still at 1.4Ghz or so if they increased it bigtime. >3. Don't make decisions based on the number of functional units. In reality, >those are never full - they are limited by various dependencies in the code. >*ALPHA 21264* with a real ISA and a very aggressive ISA doesn't even hit 2 IPC >when run on Spec Int 2k. Hammer has the same number of FUs as the Athlon. >Increasing the number of FUs does not give you infinite performance increases. You just mention a few processors (alpha 21264) which have huge penalties for branch mispredictions. A processor that can't do on paper already more than 2 instructions a clock, of course will never be able to do more than 2. So you definitely need a shitload of instructions a clock, in order to be *enabled* to do that many a clock. Only then we can look further to the cpu's weakest spot. 1 decoder at the P4 is really pathetic. I am no big hardware expert, so i don't know why they did this. Perhaps to get the chip cheaper and clocked higher? >anthony
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.