Question to Peter Martan

[peter]

Thanks. But this is one thread bench. Correct "bench 128 16/32 19" in console. Still surprising, that HT on is faster.

Maybe this time

Code: Select all

HT off

Stockfish dev-20231202-08cdbca5 by the Stockfish developers (see AUTHORS file)
bench 8192 16 27
===========================
Total time (ms) : 122165
Nodes searched  : 2251519516
Nodes/second    : 18430151


Stockfish dev-20231202-08cdbca5 by the Stockfish developers (see AUTHORS file)
bench 4098 8 25
===========================
Total time (ms) : 69076
Nodes searched  : 738661470
Nodes/second    : 10693460

HT on

Stockfish dev-20231202-08cdbca5 by the Stockfish developers (see AUTHORS file)
bench 8192 32 27
===========================
Total time (ms) : 188477
Nodes searched  : 4077313883
Nodes/second    : 21632951

Stockfish dev-20231202-08cdbca5 by the Stockfish developers (see AUTHORS file)
bench 4096 16 25
===========================
Total time (ms) : 63934
Nodes searched  : 1136924555
Nodes/second    : 17782784

Yet without any mistypings I again didn't fully make it, at trial HT off with 8 threads I wrote 4098 instead of 4096 (Mb hash), pity didn't see it before having copy- pasted and alreday rebooted, didn't want to repeat because of that, guess it won't matter much
And two more details: in BIOS of my Ryzen 5950 I can choose between SMT Disabled and Auto, Enabled isn't any option. In Task Manager with HT off and 16 threads I see 100% CPU- usage, with 32 threads HT on I see 93%, regards

[Jouni]

Very informative. Net is full of BULLSHIT that HT gives +30% to +50% speedup (also Programming and Technical Discussions). Here HT gives +17% in nps. But note: bench time increases from 122 s to 188 s. So -54% slowdown .

[peter]

Very informative. Net is full of BULLSHIT that HT gives +30% to +50% speedup (also Programming and Technical Discussions). Here HT gives +17% in nps. But note: bench time increases from 122 s to 188 s. So -54% slowdown .

Yet you saw on the other hand the number of nodes searched were almost double (x1.8) at 32 threads HT on compared to those with 16 HT off too, didn't you?
16t HT off
Nodes searched : 2 251 519 516
32t HT on
Nodes searched : 4 077 313 883

[Jouni]

With more threads SF has less reductions e.g. more careful search. I expect HT on/off has zero measurable elo difference. There was nice pre-NNUE test by fastgm http://www.talkchess.com/forum3/viewtop ... 68#p848868. He has retired?

[hgm]

I don't own any PC with HT . Hyperthreading means running 2 prosesses alternately with one CPU. Speedup is sure 0% says common sense.

No, that is not what it means at all.