Stockfish node speed mystery

[Jouni]

I tested 1 core vs 6 core with this result:

Code: Select all

Games Completed = 100 of 100 (Avg game length = 86.408 sec)
Settings = Gauntlet/640MB/30000ms+300ms/M 450cp for 5 moves, D 100 moves/PGN:Hert500.pgn(500)
Time = 8738 sec elapsed, 0 sec remaining
 1.  SFdev 6th                	63.0/100	26-0-74  	(L: m=0 t=0 i=0 a=0)	(D: r=43 i=12 f=1 s=0 a=18)	(tpm=688.8 d=32.62 nps=14139869)
 2.  SFdev 1th                	37.0/100	0-26-74  	(L: m=0 t=0 i=0 a=26)	(D: r=43 i=12 f=1 s=0 a=18)	(tpm=699.3 d=27.01 nps=2293977)

Do You see something weird? Average node speed with 6 cores in more than 6x single core speed. Ration is 6,16 . Really good scaling with NNUE. With classic evaluation ration is 5,7. Any explanation?

[Ras]

Any explanation?

Maybe something with the hardware, frequency scaling, and thermals. Perhaps like this: six cores running, CPU throttling down. One core running, CPU still in the thermal limit, but cooling off. Six cores again, initially enjoying a cooler CPU, ramping up until it hits the thermals again. All of that with 86s game length on average, so the average thinking time would be rather fast so that a delayed thermal effect could even play a role.

[Jouni]

Thre is CPU throttling with 6 cores and NNUE. But I don't understand why just 1 core is slower than 6 core. Same test with 4 cores:

Code: Select all

 1.  SFdev 1th                   	8.0/18	0-2-16  	(L: m=0 t=0 i=0 a=2)	(D: r=9 i=3 f=1 s=0 a=3)	(tpm=671.7 d=28.16 nps=2204722)
 2.  SFdev 4th                   	10.0/18	2-0-16  	(L: m=0 t=0 i=0 a=0)	(D: r=9 i=3 f=1 s=0 a=3)	(tpm=686.6 d=32.18 nps=9238251)

Speed ration now 4,19. Maybe LittleBlitzer or SF count node speed wrongly?

[Raphexon]

Or maybe hybrid eval gets called more with more cores.

[syzygy]

Or maybe hybrid eval gets called more with more cores.

Or maybe there is less pruning with more cores. Less pruning increases nps.

[Ras]

You could rule out or confirm the previous two postings by starting 1T SF four times in parallel instead of comparing to one instance of 4T SF.

[Dann Corbit]

If hash hits are counted as nodes, then it is explained easily.

[acepoint_de]

Do You see something weird? Average node speed with 6 cores in more than 6x single core speed. Ration is 6,16 . Really good scaling with NNUE. With classic evaluation ration is 5,7. Any explanation?

When doing benchmarks on my M1 I experienced the same and was laughed at. But I didn't investigate further because regarding the benchmarks I don't see it as that important: https://acepoint.de/benchmarks-with-mor ... -apple-m1/

Ciao

acepoint