Author: Robert Hyatt
Date: 08:38:28 02/19/03
Go up one level in this thread
On February 18, 2003 at 23:39:18, Aaron Gordon wrote:
>Hyatts box benchmarks a little over 2 million nodes/sec in the crafty benchmark.
>Hyatt, I know you say you get 2.5 million but in all the games you've played
>online where you've kibbed your nodes/sec you've ALWAYS been just over 2 million
>(2050-2150kn/s most of the time). Also, about a dual P4-3.06 + HT getting 3
>million nodes/sec. Not going to happen (Possible in the endgame with almost no
>pieces). My AthlonXP 2100+ overclocked to 2.44GHz gets a little over 1.5 million
>nodes/sec in crafty and I've seen over 2 million in end-games. Enricos 2400+ @
>2.52GHz gets almost 1.6 million in crafty.
Here are some numbers from a game just played:
time=3:47 cpu=398% mat=-2 n=532216437 fh=93% nps=2343k
time=4:25 cpu=399% mat=-2 n=615322858 fh=92% nps=2316k
time=2:52 cpu=399% mat=-2 n=381723163 fh=93% nps=2215k
time=3:12 cpu=398% mat=-2 n=432413276 fh=92% nps=2243k
time=3:39 cpu=398% mat=-2 n=509021660 fh=90% nps=2317k
time=2:01 cpu=399% mat=-2 n=280114908 fh=90% nps=2303k
time=1:49 cpu=399% mat=-2 n=248735781 fh=90% nps=2273k
time=2:53 cpu=398% mat=-3 n=403103259 fh=90% nps=2324k
time=2:52 cpu=399% mat=-3 n=400323062 fh=90% nps=2321k
time=1:49 cpu=399% mat=0 n=259262534 fh=90% nps=2373k
time=1:48 cpu=398% mat=0 n=268405707 fh=91% nps=2479k
time=4:20 cpu=398% mat=-1 n=625737329 fh=91% nps=2405k
time=2:15 cpu=397% mat=-1 n=333429338 fh=92% nps=2457k
time=1:47 cpu=398% mat=-1 n=259890267 fh=92% nps=2418k
time=1:24 cpu=397% mat=0 n=204315540 fh=94% nps=2408k
time=1:29 cpu=396% mat=-3 n=218637842 fh=94% nps=2454k
time=1:19 cpu=398% mat=-1 n=191197121 fh=93% nps=2415k
time=1:17 cpu=398% mat=2 n=185866947 fh=92% nps=2404k
time=1:49 cpu=398% mat=1 n=266422638 fh=93% nps=2434k
time=1:12 cpu=397% mat=1 n=176503700 fh=93% nps=2427k
time=1:10 cpu=398% mat=1 n=167575005 fh=93% nps=2360k
time=1:09 cpu=398% mat=1 n=171500968 fh=95% nps=2479k
time=1:44 cpu=398% mat=1 n=254712457 fh=94% nps=2429k
time=1:05 cpu=396% mat=1 n=159894553 fh=94% nps=2448k
time=57.20 cpu=397% mat=1 n=135742820 fh=94% nps=2373k
time=1:02 cpu=398% mat=1 n=147915710 fh=94% nps=2374k
time=3:49 cpu=397% mat=1 n=593274205 fh=97% nps=2581k
time=2:16 cpu=397% mat=1 n=351235842 fh=94% nps=2565k
time=3:59 cpu=398% mat=1 n=608866251 fh=94% nps=2545k
time=1:04 cpu=398% mat=1 n=165807816 fh=94% nps=2575k
time=17.53 cpu=395% mat=-4 n=44183665 fh=95% nps=2520k
time=1:20 cpu=394% mat=0 n=195342773 fh=93% nps=2434k
So, there _are_ a bunch of 2.5M numbers in there. Before castling, crafty is
almost
always under 2M, after castling, it varies. 2.5M is less frequent when I have
endgame
tables turned on as I always do, but if I turn them off in endgames, it can
reach beyond
3.0M.
As I have said _often_ there is no one NPS value that is correct. I usually
refer to 2.5M
as the typical number in the middlegame, which is fairly close. If you want to
argue it
should be 2.4M, you have a point. If you want to argue it should be 2.0 because
of the
opening, that is also a point. But in middlegame positions, such as those in
test suites,
Here are a couple of outputs:
WAC:
average search depth.............. 4.9
nodes per second.................. 2408280
ECM (subset):
average search depth.............. 7.3
nodes per second.................. 2521680
So the numbers make reasonable sense to talk about 2.4M-2.5M.
As far as your overclocking stuff goes, feel free to do it. I take the results
with a
grain of salt, because I understand things like "settling time". I once spent
two weeks
debugging a problem on an AMD machine that was overclocked (I did not know it at
the time). I then discovered the program ran _perfectly_ on an intel box I had,
but would
fail after several hours on the AMD. I reset everything to factory specs and
the program
ran fine on the AMD as well.
Circuits have a distinct "settling time" that varies depending on inputs and
other things. If
you cut the clock cycle time so that this circuit settles almost all the time,
before the clock
edge falls, then you will get right answers almost all the time. But there are
those rare
exceptions where bogus results pop out, and I don't want them in _my_ program.
I don't
care _how_ well it is tested, unless you do an _exhaustive test_ (impossible)
there is always
room for doubt...
I want solid results...
Feel free to use any NPS number you want. My 2.5M is pretty close to normal for
most of the game, and it can go up above that without 200 gigs of 6 man egtb
tables.
>
>There are *NO* P4 chips available that can match this, including a P4-3.06 Xeon
>+ HT. Those get around 1.3 million WITH HT in crafty. If you do a 1.3 million *
>1.8 you come out to 2.34 million nodes/sec. Far cry from 3 million nodes/sec.
>
>How much did MY box cost? $97 for the chip, $80 for the board. $75 for the ram,
>etc. Not much at all.
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.