Komodo vs. SF scaling

[cma6]

I've been testing infinite analysis of Komodo 10.1 in Aquarium on a dual-Xeon (36) core machine with 64 Gb of RAM. In my tests the "threads" parameter in the engine = # of actual cores used.
My preliminary tests on a typical middle-game position show that SF (Haswell version) scales much better: SF with 35 cores 31,867 Kn/s vs. 10,511 Kn/s with 6 cores.
Komodo 10.1 did not run properly with 35 threads (=cores);
with 24 cores, 25,225 Kn/s vs. 8688 Kn/s with 6 cores.
How have others done with Komodo scaling?
And is there not a Komodo parameter setting that is supposed to help with better interaction with the GUI?

[Tdunbug]

I concur with your results. I have played with both SF and K 10.1 on my 20/40 ht machine and K preforms a little worse. I don't know the answer to this.

[bob]

I've been testing infinite analysis of Komodo 10.1 in Aquarium on a dual-Xeon (36) core machine with 64 Gb of RAM. In my tests the "threads" parameter in the engine = # of actual cores used.
My preliminary tests on a typical middle-game position show that SF (Haswell version) scales much better: SF with 35 cores 31,867 Kn/s vs. 10,511 Kn/s with 6 cores.
Komodo 10.1 did not run properly with 35 threads (=cores);
with 24 cores, 25,225 Kn/s vs. 8688 Kn/s with 6 cores.
How have others done with Komodo scaling?
And is there not a Komodo parameter setting that is supposed to help with better interaction with the GUI?

Wrong way to measure scaling. It is not about NPS, it is about how quickly you search to a deeper depth.

[Evert]

I've been testing infinite analysis of Komodo 10.1 in Aquarium on a dual-Xeon (36) core machine with 64 Gb of RAM. In my tests the "threads" parameter in the engine = # of actual cores used.
My preliminary tests on a typical middle-game position show that SF (Haswell version) scales much better: SF with 35 cores 31,867 Kn/s vs. 10,511 Kn/s with 6 cores.
Komodo 10.1 did not run properly with 35 threads (=cores);
with 24 cores, 25,225 Kn/s vs. 8688 Kn/s with 6 cores.

I don't see how SF "scales better": you report that NPS increases by a factor of 3 if you increase the number of cores by a factor of 6. For Komodo you report an increase in NPS by a factor of 3 for 4x the number of cores. Seems to me Komodo scales better; what am I missing?

Note that you can get perfect NPS scaling by just having N threats search the same position without communicating, so it's not a particularly useful statistic by itself.