Page 3 of 9
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 11:54 pm
by dragontamer5788
zullil wrote: ↑Mon Aug 12, 2019 11:19 pm
dragontamer5788 wrote: ↑Mon Aug 12, 2019 11:14 pm
Or are all of the cores
unnecessarily searching nodes because they don't have any idea of what to actually do?
Nice description of "Lazy SMP".
Well... the description unfortunately applies to ABDADA and YBWC as well. Which is why all of my chess work is entirely focused on trying to discover a work-efficient search methodology.
I hoped that my current methodology would be 100% work efficient, but alas... a fully work efficient methodology would have limited parallelism (average case speedup of fractions of a percent), and require an exponential amount of RAM to store the huge amount of parallel processing (roughly O(ply^(depth/2)) storage costs). Needless to say, that's unworkable as ply-searched grows. Again,
MY method has this problem, maybe someone else's future discovery can figure out a way around these problems.
Still, I think I've discovered something that is
more work efficient than current methodologies. So I'm willing to code it up and test it. I think I've come up with a better speculative execution heuristic than YBWC or ABDADA.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Tue Jan 14, 2020 7:21 pm
by yorkman
Were those SF and asmFish results running with HT enabled?
And there's several benchmarks that shows these same engines on a dual 7742 get only about 196 Mn/s:
https://openbenchmarking.org/result/191 ... XAMDEPYC12
Then there's STH's benchmark on a different system (Daytona_X) and got 275 Mn/s for asmFish which was with HT on (about 150 Mn/s on one 7742):
http://ipmanchess.yolasite.com/amd---in ... -bench.php
Either way, these scores are quite conflicting. So if I planned to buy 2x 7742 cpu's what speeds can I expect in SF dev, especially in Windows? Will I get 190 Mn/s? 230 Mn/s? 275 Mn/s?
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Wed Jan 15, 2020 1:54 am
by Leo
Werewolf wrote: ↑Sun Aug 11, 2019 7:59 pm
Surpassing Deep Blue's 200 Million nps.
But 22 years later...
I think Deep Blue used brute force so it needed a huge NPS to search. I dont think its pruning was good.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Wed Jan 15, 2020 5:11 am
by MikeB
yorkman wrote: ↑Tue Jan 14, 2020 7:21 pm
Were those SF and asmFish results running with HT enabled?
And there's several benchmarks that shows these same engines on a dual 7742 get only about 196 Mn/s:
https://openbenchmarking.org/result/191 ... XAMDEPYC12
Then there's STH's benchmark on a different system (Daytona_X) and got 275 Mn/s for asmFish which was with HT on (about 150 Mn/s on one 7742):
http://ipmanchess.yolasite.com/amd---in ... -bench.php
Either way, these scores are quite conflicting. So if I planned to buy 2x 7742 cpu's what speeds can I expect in SF dev, especially in Windows? Will I get 190 Mn/s? 230 Mn/s? 275 Mn/s?
The high end results were all on Linux and my guess is that Linux Huge pAges were enabled - which make a big big difference.
On a Linux box with a Threadripper 3970x 32C- ( btw the entire computer incljuding RTX 2060 Super with 128 GB ram was 33% less then just one 7742 CPU - getting 2x 7742 is a lot of cheese to throw on something just for chess 8>0 ..)
I think the Patrick @ ServeTheHome.com results on Ipman's web page are pretty close to what you should get on a Linux box, maybe subtract up to about 10% for a Windows box - I really don't know since I don't use Windows.
Code: Select all
my normal asmfish bench (huge pages are enabled)
*** bench hash 1024 threads 64 depth 26 realtime 0 ***
info string hash set to 1024 MB no large pages
info string node 0 has threads 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63
1: nodes: 1126178343 100184 knps
2: nodes: 1460651697 114158 knps
3: nodes: 50390197 143153 knps
4: nodes: 429945418 123299 knps
5: nodes: 407909363 110844 knps
6: nodes: 423555728 109361 knps
7: nodes: 768386305 113700 knps
8: nodes: 1832037588 117093 knps
9: nodes: 1026478155 110922 knps
10: nodes: 168454412 123500 knps
11: nodes: 1134875669 110064 knps
12: nodes: 560759868 107343 knps
13: nodes: 157923327 118828 knps
14: nodes: 2181008929 110576 knps
15: nodes: 308828275 120919 knps
16: nodes: 174292914 143806 knps
17: nodes: 113405334 147663 knps
18: nodes: 185294367 141122 knps
19: nodes: 56795017 149068 knps
20: nodes: 236038061 157884 knps
21: nodes: 42374105 128018 knps
22: nodes: 41036434 112428 knps
23: nodes: 414670048 135557 knps
24: nodes: 245236234 139259 knps
25: nodes: 273591967 170144 knps
26: nodes: 8325162 136478 knps
27: nodes: 20546163 129221 knps
28: nodes: 251710465 139143 knps
29: nodes: 89787424 126818 knps
30: nodes: 27209330 138118 knps
31: nodes: 13012135 110272 knps
32: nodes: 5225319 102457 knps
33: nodes: 3380388 80485 knps
34: nodes: 25931781 131633 knps
35: nodes: 17650320 122571 knps
36: nodes: 4090462 107643 knps
37: nodes: 4172004 126424 knps
===========================
Total time (ms) : 123442
Nodes searched : 14291158708
Nodes/second : 115772255
with 32 cores , no HT 0 about 75 Mnps
for cur-dev-SF, huge pages enabled
HT on, 64 threads
bench 1024 64 26 ( the Ipman bench)
Position: 1/46
Nodes/Second: 95536k
Position: 2/46
Nodes/Second: 102022k
Position: 3/46
Nodes/Second: 158897k
Position: 4/46
Nodes/Second: 114007k
Position: 5/46
Nodes/Second: 101610k
Position: 6/46
Nodes/Second: 104037k
Position: 7/46
Nodes/Second: 106838k
Position: 8/46
Nodes/Second: 110684k
Position: 9/46
Nodes/Second: 99727k
Position: 10/46
Nodes/Second: 116862k
Position: 11/46
Nodes/Second: 99118k
Position: 12/46
Nodes/Second: 98567k
Position: 13/46
Nodes/Second: 113381k
Position: 14/46
Nodes/Second: 101118k
Position: 15/46
Nodes/Second: 117711k
Position: 16/46
Nodes/Second: 135196k
Position: 17/46
Nodes/Second: 154340k
Position: 18/46
Nodes/Second: 180698k
Position: 19/46
Nodes/Second: 178519k
Position: 20/46
Nodes/Second: 163144k
Position: 21/46
Nodes/Second: 221856k
Position: 22/46
Nodes/Second: 219120k
Position: 23/46
Nodes/Second: 231569k
Position: 24/46
Nodes/Second: 148659k
Position: 25/46
Nodes/Second: 200063k
Position: 26/46
Nodes/Second: 148403k
Position: 27/46
Nodes/Second: 152118k
Position: 28/46
Nodes/Second: 129176k
Position: 29/46
Nodes/Second: 125304k
Position: 30/46
Nodes/Second: 171878k
Position: 31/46
Nodes/Second: 119905k
Position: 32/46
Nodes/Second: 105741k
Position: 33/46
Nodes/Second: 102924k
Position: 34/46
Nodes/Second: 104017k
Position: 35/46
Nodes/Second: 185218k
Position: 36/46
Nodes/Second: 179839k
Position: 37/46
Nodes/Second: 168020k
Position: 38/46
Nodes/Second: 159001k
Position: 39/46
Nodes/Second: 164089k
Position: 40/46
Nodes/Second: 164985k
Position: 41/46
Nodes/Second: 144960k
Position: 42/46
Nodes/Second: 93949k
Position: 43/46
Nodes/Second: 78843k
Position: 44/46
Nodes/Second: 0k
Position: 45/46
Nodes/Second: 0k
Position: 46/46
Nodes/Second: 92109k
===========================
Total time (ms) : 69607
Nodes searched : 7705105461
Nodes/second : 110694405
with HT, just using the 32 real cores
Position: 1/46
Nodes/Second: 60533k
Position: 2/46
Nodes/Second: 65475k
Position: 3/46
Nodes/Second: 101888k
Position: 4/46
Nodes/Second: 72472k
Position: 5/46
Nodes/Second: 65218k
Position: 6/46
Nodes/Second: 66042k
Position: 7/46
Nodes/Second: 69106k
Position: 8/46
Nodes/Second: 70464k
Position: 9/46
Nodes/Second: 64712k
Position: 10/46
Nodes/Second: 74214k
Position: 11/46
Nodes/Second: 63111k
Position: 12/46
Nodes/Second: 63536k
Position: 13/46
Nodes/Second: 71324k
Position: 14/46
Nodes/Second: 64424k
Position: 15/46
Nodes/Second: 74207k
Position: 16/46
Nodes/Second: 73934k
Position: 17/46
Nodes/Second: 86595k
Position: 18/46
Nodes/Second: 106470k
Position: 19/46
Nodes/Second: 111233k
Position: 20/46
Nodes/Second: 101423k
Position: 21/46
Nodes/Second: 137323k
Position: 22/46
Nodes/Second: 136304k
Position: 23/46
Nodes/Second: 147236k
Position: 24/46
Nodes/Second: 94001k
Position: 25/46
Nodes/Second: 122131k
Position: 26/46
Nodes/Second: 92183k
Position: 27/46
Nodes/Second: 93389k
Position: 28/46
Nodes/Second: 81633k
Position: 29/46
Nodes/Second: 78957k
Position: 30/46
Nodes/Second: 100512k
Position: 31/46
Nodes/Second: 75417k
Position: 32/46
Nodes/Second: 67675k
Position: 33/46
Nodes/Second: 65501k
Position: 34/46
Nodes/Second: 66488k
Position: 35/46
Nodes/Second: 113165k
Position: 36/46
Nodes/Second: 113417k
Position: 37/46
Nodes/Second: 102312k
Position: 38/46
Nodes/Second: 102213k
Position: 39/46
Nodes/Second: 104582k
Position: 40/46
Nodes/Second: 101449k
Position: 41/46
Nodes/Second: 93789k
Position: 42/46
Nodes/Second: 45257k
Position: 43/46
Nodes/Second: 61276k
Position: 44/46
Nodes/Second: 0k
Position: 45/46
Nodes/Second: 0k
Position: 46/46
Nodes/Second: 59805k
===========================
Total time (ms) : 66669
Nodes searched : 4756519605
Nodes/second : 71345296
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Wed Jan 15, 2020 6:17 pm
by bob
Vinvin wrote: ↑Mon Aug 12, 2019 7:50 am
Laskos wrote: ↑Mon Aug 12, 2019 7:10 am
Raphexon wrote: ↑Mon Aug 12, 2019 6:25 am
Werewolf wrote: ↑Sun Aug 11, 2019 9:53 pm
mehmet karaman wrote: ↑Sun Aug 11, 2019 8:59 pm
Vinvin wrote: ↑Sun Aug 11, 2019 8:06 pm
Werewolf wrote: ↑Sun Aug 11, 2019 7:59 pm
Surpassing Deep Blue's 200 Million nps.
But 22 years later...
But now, Stockfish at 1 Mn/s is stronger than Deep Blue at 200 Mn/s.
Stockfish at 1 kn/s is stronger than Deep Blue.
Is it? I’m not sure that’s right
Stockfish should be GM level even at 100 n/s classical time control.
Probably 0.5 knps at classical time control, if my old inferrence works. Super GM level at some 5 knps.
I often say that Stockfish reach 2600-2700 Elo with 1 second by move on an average home machine (around 2 kn/s) or a fast phone against human on a long game (40 moves in 2h). This view set SF speed around 10 kn/s.
I do not believe this is correct. At 5K nodes per second, you get to search 5K nodes and you are done. Not much of a search. I have the ability in Crafty to search a specific number of nodes (approximately). For a simple test, I took the newer was positions (201 total) and ran them against standard Crafty at 1 second per move on fairly good hardware (60M nodes per second). It correctly solved all 201 positions. I then re-ran the same test, except telling crafty to stop after searching 5K nodes per second (5K nodes total here) it only solved 114 out of the 201 positions. That hardly sounds like a GM to me. If you stretch it to 60 seconds, normal Crafty still gets all 201. Crafty searching 300K nodes (60 seconds x 5K nodes per second) gets 171 right.
That 5K nodes per second is a REAL restriction. Many were doing 5K nodes per second in the 70's and 80's. And far beyond. Without a GM being produced.
I am certain that the 2700 Elo at 2K nodes per second is a wild exaggeration of reality. Maybe 500K nodes per second, possible. Certainly not 2K.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Wed Jan 15, 2020 11:39 pm
by MikeB
bob wrote: ↑Wed Jan 15, 2020 6:17 pm
Vinvin wrote: ↑Mon Aug 12, 2019 7:50 am
Laskos wrote: ↑Mon Aug 12, 2019 7:10 am
Raphexon wrote: ↑Mon Aug 12, 2019 6:25 am
Werewolf wrote: ↑Sun Aug 11, 2019 9:53 pm
mehmet karaman wrote: ↑Sun Aug 11, 2019 8:59 pm
Vinvin wrote: ↑Sun Aug 11, 2019 8:06 pm
Werewolf wrote: ↑Sun Aug 11, 2019 7:59 pm
Surpassing Deep Blue's 200 Million nps.
But 22 years later...
But now, Stockfish at 1 Mn/s is stronger than Deep Blue at 200 Mn/s.
Stockfish at 1 kn/s is stronger than Deep Blue.
Is it? I’m not sure that’s right
Stockfish should be GM level even at 100 n/s classical time control.
Probably 0.5 knps at classical time control, if my old inferrence works. Super GM level at some 5 knps.
I often say that Stockfish reach 2600-2700 Elo with 1 second by move on an average home machine (around 2 kn/s) or a fast phone against human on a long game (40 moves in 2h). This view set SF speed around 10 kn/s.
I do not believe this is correct. At 5K nodes per second, you get to search 5K nodes and you are done. Not much of a search. I have the ability in Crafty to search a specific number of nodes (approximately). For a simple test, I took the newer was positions (201 total) and ran them against standard Crafty at 1 second per move on fairly good hardware (60M nodes per second). It correctly solved all 201 positions. I then re-ran the same test, except telling crafty to stop after searching 5K nodes per second (5K nodes total here) it only solved 114 out of the 201 positions. That hardly sounds like a GM to me. If you stretch it to 60 seconds, normal Crafty still gets all 201. Crafty searching 300K nodes (60 seconds x 5K nodes per second) gets 171 right.
That 5K nodes per second is a REAL restriction. Many were doing 5K nodes per second in the 70's and 80's. And far beyond. Without a GM being produced.
I am certain that the 2700 Elo at 2K nodes per second is a wild exaggeration of reality. Maybe 500K nodes per second, possible. Certainly not 2K.
All my tests show that SF at 50K/sec - might be somewhere at the GM level - whether it is a Super GM or a more ordinary GM, who knows. I would be interested in hearing from Larry Kaufman on this. The Revelation is a device that reportedly plays at GM level and it see around 100K/nps ( roughly - I do not own one) . Perhaps only a GM knows for sure.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Wed Jan 15, 2020 11:42 pm
by Ovyron
"GM level" means that after a match is played we expect a 50% performance, I doubt any human in the world has any chance of getting close to that at 50K/sec.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Thu Jan 16, 2020 12:28 am
by MikeB
Ovyron wrote: ↑Wed Jan 15, 2020 11:42 pm
"GM level" means that after a match is played we expect a 50% performance, I doubt any human in the world has any chance of getting close to that at 50K/sec.
Why are you always so angry in your posts? I have no desire to interact with you at all. You definition of GM level is not quite accurate - there is wide range of GMs of level - from the very highest level, to those that are much weaker. Plus my words are couched - they were deliberately couched for people like you. In reality you should be banned from this site, but I guess there is no rule against someone who enjoys being unpleasant to others.
'' ....
might be somewhere at the GM level - whether it is a Super GM or a more ordinary GM, who knows.
I would be interested in hearing from Larry Kaufman on this. " did I write something in any form that was absolute? " ...might be ..." "would like to hear from Larry K" etc.
I'm sorry , but for my own sanity , i have to block you from me. Have a nice life.
さようなら
Sayōnara
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Thu Jan 16, 2020 1:16 am
by Alayan
You're overreacting, Mike. I don't think you would have interpreted his post that way if you had not seen his username.
He said that he doubt that any human in the world has any chance of getting a 50% perf against SF at 50Knps. That's not an angry statement, that's the obvious truth.
This talk about constraining SF to a low nps is a contrived way to talk about time odds (the only point of slowing down the engine is to make the human ponder). Now, the reference nps depends on hardware, but taking fishtest's 1.6mnps standard (most recent CPUs do quite better on a single-core),
50knps is about 1:30 time odds. In TCEC QL testing,
Stockfish with 1:300 time odds had a big positive score against what are 2800+ 1CPU CCRL engines. Sure, it wasn't on a single-core, but it gives a rough idea of how the massive strength gap can more than overcome the time odds.
So Stockfish with 3 minutes + 1s/move on the clock on one core would easily crush any human having a few hours on the clock.
We know that human are relatively worse at blitz against engines than at classical TC, so while with a sufficiently short TC, you'd get Stockfish on such low node counts that its strength would plumet, the human wouldn't be able to compete at all.
I mean, at bullet 1m+0s SF level 1 on lichess can give trouble to an IM that isn't careful :
https://www.youtube.com/watch?v=5FW2gfzYTiA
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Thu Jan 16, 2020 2:16 am
by MikeB
Alayan wrote: ↑Thu Jan 16, 2020 1:16 am
You're overreacting, Mike. I don't think you would have interpreted his post that way if you had not seen his username.
He said that he doubt that any human in the world has any chance of getting a 50% perf against SF at 50Knps. That's not an angry statement, that's the obvious truth.
This talk about constraining SF to a low nps is a contrived way to talk about time odds (the only point of slowing down the engine is to make the human ponder). Now, the reference nps depends on hardware, but taking fishtest's 1.6mnps standard (most recent CPUs do quite better on a single-core),
50knps is about 1:30 time odds. In TCEC QL testing,
Stockfish with 1:300 time odds had a big positive score against what are 2800+ 1CPU CCRL engines. Sure, it wasn't on a single-core, but it gives a rough idea of how the massive strength gap can more than overcome the time odds.
So Stockfish with 3 minutes + 1s/move on the clock on one core would easily crush any human having a few hours on the clock.
We know that human are relatively worse at blitz against engines than at classical TC, so while with a sufficiently short TC, you'd get Stockfish on such low node counts that its strength would plumet, the human wouldn't be able to compete at all.
I mean, at bullet 1m+0s SF level 1 on lichess can give trouble to an IM that isn't careful :
https://www.youtube.com/watch?v=5FW2gfzYTiA
Perhaps , and if he posted what he said without quoting me, there would been zero reaction from me for sure.
Your comment is interesting regarding the 1/300 time odds to say the least. i was not aware of those tests. Also, I may have a natural bias to understate the strength of Honey for no other reason not to claim something that may not be true. Anyway - it is true, that SF at very low nps , 1000 nodes per second or is is very, very strong.