237 Mn/s for Stockfish on an 2xEPYC 7742

[dragontamer5788]

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.

[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?

[Leo]

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.

[MikeB]

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

[bob]

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.

[MikeB]

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.

[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.

[MikeB]

"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

[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

[MikeB]

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.