Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 9:10 am
Probably 0.5 knps at classical time control, if my old inferrence works. Super GM level at some 5 knps.
Page 2 of 9
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 9:50 am
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.Laskos wrote: ↑Mon Aug 12, 2019 9:10 amProbably 0.5 knps at classical time control, if my old inferrence works. Super GM level at some 5 knps.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 9:51 am
I'm in ... saying my pennies now.Vinvin wrote: ↑Sun Aug 11, 2019 9:52 pm The new EPYC 7742 runs Stockfish quite fast !
The EPYC 7502 is new as well.
All the results here : https://www.phoronix.com/scan.php?page= ... 7742&num=5
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 10:49 am
also, just for reference, first microprocessor chess programs, Z80, 6502, back in 1980, 1981 were doing around 300 nps.
Factor of about a million in 40 years by general purpose CPU/software combination.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 10:53 am
Elo performance of Pocket Fritz 4 (Hiarcs 13 Engine) is 2938 elo vs. GM's and IM's at 2009 Mercosur CupWerewolf wrote: ↑Sun Aug 11, 2019 11:53 pm
Is it? I’m not sure that’s right
Only 1 core, 528 mHz HTC Touch HD mobile phone.
And it searches fewer than 20 kn/s positions per second.
https://en.chessbase.com/post/breakthro ... enos-aires
Hiarcs 13.1 elo is 2874 (CCRL 40/40, 1 cpu) , Stocfish 10 elo is 3387 (CCRL 40/40, 1 cpu)
Elo difference is 513
Hiarcs 13.1 elo is 2728 (CEGT 20/20, 1 cpu), Stockfish 10 elo is 3419 (CEGT 20/20, 1cpu)
Elo difference is 691
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 11:34 am
Yes, maybe you are right !mehmet karaman wrote: ↑Mon Aug 12, 2019 10:53 amElo performance of Pocket Fritz 4 (Hiarcs 13 Engine) is 2938 elo vs. GM's and IM's at 2009 Mercosur CupWerewolf wrote: ↑Sun Aug 11, 2019 11:53 pm
Is it? I’m not sure that’s right
Only 1 core, 528 mHz HTC Touch HD mobile phone.
And it searches fewer than 20 kn/s positions per second.
https://en.chessbase.com/post/breakthro ... enos-aires
Hiarcs 13.1 elo is 2874 (CCRL 40/40, 1 cpu) , Stocfish 10 elo is 3387 (CCRL 40/40, 1 cpu)
Elo difference is 513
Hiarcs 13.1 elo is 2728 (CEGT 20/20, 1 cpu), Stockfish 10 elo is 3419 (CEGT 20/20, 1cpu)
Elo difference is 691
1 kn/s is more than 1/1000 time handicap !
7 seconds for Stockfish (on a slow computer) and 2 hours for its opponent ! Very funny !
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Mon Aug 12, 2019 11:36 am
My oldest functioning CPU is Pentium E5200 which is obviously to new and strong at single core.
Would be fun to know how 15-20 year old CPUs do in 2019.
Alternatively Chessmaster 8 or 9 at single core.
Would be fun to know how 15-20 year old CPUs do in 2019.
Alternatively Chessmaster 8 or 9 at single core.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Tue Aug 13, 2019 12:58 am
It looks like 64 cores would cost $5,000.
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Tue Aug 13, 2019 1:14 am
Mn/s is obviously the measurement that scales best with more CPU power. So as a CPU-benchmark is concerned its the best stat to discuss.
Unfortunately, the best stat for chess programmers and chess players to discuss is Time-to-depth: all current Alpha-Beta search algorithms have a degree of "search overhead waste", where different threads may unnecessarily search some nodes. They speculatively execute nodes because the alpha/beta bounds are unknown.
As such, Chess Programmers and chess users need to measure Time-to-depth. Do these 64 cores or 128 cores actually speed up a Ply15 search? Or are all of the cores unnecessarily searching nodes because they don't have any idea of what to actually do?
Unfortunately, the best stat for chess programmers and chess players to discuss is Time-to-depth: all current Alpha-Beta search algorithms have a degree of "search overhead waste", where different threads may unnecessarily search some nodes. They speculatively execute nodes because the alpha/beta bounds are unknown.
As such, Chess Programmers and chess users need to measure Time-to-depth. Do these 64 cores or 128 cores actually speed up a Ply15 search? Or are all of the cores unnecessarily searching nodes because they don't have any idea of what to actually do?
Re: 237 Mn/s for Stockfish on an 2xEPYC 7742
Posted: Tue Aug 13, 2019 1:19 am
Nice description of "Lazy SMP".dragontamer5788 wrote: ↑Tue Aug 13, 2019 1:14 am Or are all of the cores unnecessarily searching nodes because they don't have any idea of what to actually do?
