Partial Multi-Threaded Performance of AMD and Intel Chips

[CornfedForever]

Multi-threading question about Cores/Threads and NPS for Stockfish (or other engines) concerning the top consumer CPU's on the market today at stock.
Every reference to NPS I see seems to pertain to throwing the full CPU at the question - all 32 threads....but what if you use less, omitting any 'efficiency core' input.

Ex:
Intel 13th Gen i913900K
8 ‘Performance’ Cores/ 16 ‘Efficiencey single thread) = 32 Threads

AMD Ryzen 7950X
16 ‘Performance’ Cores/32 Threads

I want to do other things while the engine analyses and presume all these other things being equal.

Lets say I want to use all 8 performance cores (16 threads) on the Intel chip on a position and want to use 8 cores (16 threads) on the AMD chip as well....freeing up power to do other things for either while the engine is analyzing?

What are the NPS likely to look like for each under these conditions? Would AMD still hold an edge over the Intel chip in multi-threaded engine NPS (?) or does that edge at full use completely go away in my scenario?

[smatovic]

From Ipman SF 14.1 bench:
https://ipmanchess.yolasite.com/amd--in ... ckfish.php

42.704.893 AMD Ryzen 9 7950X @4.7Ghz DDR5 6000 CL30 16cores
per core:
42M NPS /16cores = 2.625M NPS/core; x 8cores = 21M NPS (w/o HyperThreading)

31.480.909 Intel Core i9 13900 ddr5 4800 CL36 32threads (with HyperThreading)
per thread:
31.5M NPS / 32threads = ~0.98M NPS/thread

Question, how much is HT (SMT-2, two threads per core) for SF on 13th gen? Let's say it is x1.5 in NPS.

Question, how much faster is an Intel performance core compared to efficiency core?

Geekbench P-Core vs. E-Core on Anandtech:
https://www.anandtech.com/show/17047/th ... plexity/10

Let's say it is x1.66.

Let's say 8 P-cores with 16 threads do ~21M NPS and 16 E-cores do ~10M NPS.

So, "very roughly estimated": 8 Intel P-cores x 1.75M NPS = 14M NPS

Dunno how to stick engine on P-core though, dunno how to run HT in regard to E-Core (OS vs engine), Intel 12th gen on Ipman bench seems faster.

Or, alternatively, just buy an Apple

--
Srdja

[CornfedForever]

Yes, thanks, the ipman site was the first I went to, yes, and it's easy to see at full usage, the AMD outperforms the Intel on multi-threaded engine performance.
One can do the math as you did but hat still leaves me wondering.

But if one only uses 2, 4 6, 8 Performance cores for each. (no efficiency cores)...I'm wondering if just using the same number of cores/threads results in essentially the same NPS or if one chip is inherently 'better/faster' than the other (ram, motherboard, etc essentially being the same).

I simply can't find where anyone has done that kind of testing. It would be nice to know as I am shopping for a new PC or laptop. I know I would be happy with either processor...but just wondering.

[Modern Times]

But if one only uses 2, 4 6, 8 Performance cores for each. (no efficiency cores)...I'm wondering if just using the same number of cores/threads results in essentially the same NPS or if one chip is inherently 'better/faster' than the other (ram, motherboard, etc essentially being the same).

Depends exactly how important it is to you. Very likely that performance is within 10% either way. If you're not worried about every last ounce of performance, then it probably doesn't matter. What I will say though is that with the big-little Intel architecture you may need to set affinities, but depends what you are doing with the engines - casual analysis or alternatively ratings measurement ?

[CornfedForever]

But if one only uses 2, 4 6, 8 Performance cores for each. (no efficiency cores)...I'm wondering if just using the same number of cores/threads results in essentially the same NPS or if one chip is inherently 'better/faster' than the other (ram, motherboard, etc essentially being the same).

Depends exactly how important it is to you. Very likely that performance is within 10% either way. If you're not worried about every last ounce of performance, then it probably doesn't matter. What I will say though is that with the big-little Intel architecture you may need to set affinities, but depends what you are doing with the engines - casual analysis or alternatively ratings measurement ?

Thanks for the thought Modern Times.
I would not call it 'casual' analysis...pretty serious, but no, not engine vs engine nor trying to peg a rating to a given engine.

I was just curious as to 'direct matchup' of the best cores/same number for each and can't find any real comparison (just good guesses). 1v1core (2 threads) up to 8 v 8 (16 threads).

Practically, it doesn't really matter. I mean if one is 10% better than the other...for analysis you could arguably run the slower ever so very slightly longer to get the same performance.

I would say that I am also considering the 'laptop equivalent' of those CPU's and have a slight fear that the AMD 7945x gets so hot under...looks like under any currently available laptop, that one might need to lower the amount of electricity you let it eat to...maybe 80% for long terms or overnight analysis on a bunch of positions since it runs so hot. Of course it is more efficient than the intel version, but maybe one could let the Intel chip eat more wattage and be okay under the same conditions. Just tumbling things over in my head there.