Author: Robert Hyatt
Date: 06:35:11 03/04/00
Go up one level in this thread
On March 04, 2000 at 00:08:14, Albert Silver wrote: >On March 03, 2000 at 22:16:39, Robert Hyatt wrote: > >>On March 03, 2000 at 20:05:57, Tom Kerrigan wrote: >> >>>On March 03, 2000 at 17:08:58, Robert Hyatt wrote: >>> >>>>Problem is the compilers don't know what is going on. IE how many "hidden" >>>>registers does the architecture have for renaming? Intel (nor anyone else) >>>>will make this a 'constant'. >>> >>>But my point is, why have register renaming at all. I can list a dozen good >>>processors that don't do it. I would like to know the exact percentage speedup >>>it gives you. >> >> >>That is probably lost in the benchmark data. But in the case of intel, with >>8 (barely) usable registers, it would be impossible to keep multiple pipes >>busy due to register conflicts. Renaming solves this nicely and frees up >>parallel streams of instructions to keep the pipes busy... >> >>Some don't need to do this, like the sparc/ultrasparcs, because they have >>32 accessible registers for programming. But 8? What a decision... :) >> >> >> >> >>> >>>>And how does the P5 do more per cycle than a P6 when the p6 can do three >>>>ops/cycle, while the P5 drags along at a max of 2, and it requires a very >>>>good compiler to do two at a time??? >>> >>>TSCP (1.42) on an original Pentium/200 searches 136 NPS/MHz. >>> >>>On a Pentium II/300, it searches 119 NPS/MHz. >>> >>>So the Pentium appears to be 15% faster, despite its lack of out-of-order >>>execution, branch predition, speculative execution, register renaming, >>>reservation stations, blah blah blah. >> >> >>That looks like a compiler issue. I ran crafty on a P5/233mmx, and a P6/200, >>and the P5 was about 70% as fast (I have a P5/233 notebook laying around, >>and while it was not bad, it was definitely slower than a p6/200 machine >>for all the things I tested (ie linux kernel builds, crafty testing, etc.) >> >> >> >> >>> >>>I assume this is because the P5 has shorter pipes and doesn't have to flush them >>>all the time due to speculative execution gone wrong. >>> >>>(BTW, the K5 has almost everything beat. It searches 173 NPS/MHz, and it doesn't >>>do anything particularly fancy either.) >> >>same point as above. For crafty, the K* processors are slower. I have not >>delved into why... >> > >How about on the Athlons? > > A.S. > > no comparative data so far... > > >> >> >> >>> >>>-Tom
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.