Author: Vincent Diepeveen
Date: 00:13:53 01/21/02
Go up one level in this thread
On January 21, 2002 at 00:04:50, Terry McCracken wrote: >On January 20, 2002 at 23:43:21, Vincent Diepeveen wrote: > >>On January 20, 2002 at 22:02:10, Terry McCracken wrote: >> >>>Can anyone explain how and why a P-IV at 1.7 Ghz with 512MB (PC800) RDRAM >>>outperformed an Athlon running at 1.4 Ghz with 266Mgh DDR Bus and 256MB of SDRAM >>>on a Fritzmark by a wide margin? >> >>fritz might be in p3 assembly and has probably nearly no mispredicted >>branches somehow. further it is such a small program that it hardly >>needs data cache. >> >I'm not sure what you mean that Fritz is a small programme compared to other >chess software? If that is so, then why is Fritz at the head of it's class >compared to other commercial programmes most of the time wich you say need more >cache and are larger programmes? each node in fritz used to cost under 1000 clock ticks. right now it's probably a bit less than 2000 clocks a node. In short it means fritz is more busy with quantity than quality of a node. Some years ago fritz was sold with next sentence: "fritz learns through search". nowadays that is no longer seen as 'good', but i remember 1997 when some major company was making PR by just quoting number of nodes. Search depths or hashtables or search quality didn't matter. Only number of nodes a second. >>The 3 major disadvantages from the P4 are >> a) HUGE branch misprediction penalty >> b) very SMALL datacache >> c) no way to do 4 instructions a second according to 4 instructions a clock i obviously meant. >> experts who read the design, it can to at most 3 >> instructions a second. >> > >Well this was a P-4 vs Athlon, _not_ a P-3. So optmization for a P-3 would have >no bearing in this case. A misprediction at the P3 costs *at least* 10 clocks, whereas most measure it to be practical more than way above the 'on average 15 clocks' which intel gives as average for a misprediction. In 15 clocks you might execute 45 instructions. Obviously branchmisprediction at the P3 and K7 are also important. At the P4 to get a misprediction costs you: *at least* 20 clocks. Average branch misprediction penalty i didn't hear anyone talk about but it will be more near 40 than near 20. >I find much of the data about the P-4 at this site very >suspect. So it does make sense. P4 is a very new design which is released too soon. With a bit more work on the P4 it might be a killercpu, but right now it is a big waste of money for most applications. >There appeared to be no serious branch misprediction, and it flew past the >Athlon chip. a minimum of 20 clocks is like deathpenalty, really. >>> > >The Fritzmark for the Pentium was 749 and the AMD a mere 548! Why? This is >>>contrary to the findings of this forum. That is a difference of 201, that is >>>large! >>> >>>An explantion would be appreciated, thanks! >>> >>>Terry
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.