Best cpu benchmark for chess?

[towforce]

A study, seriously???
There are 3 points in this thread, there are a couple of points for M1 CPU, anyone can generate numbers for their own CPU they have. So your "study" is to collect those data do a linear regression and get a correlation coefficient.
Ofc doing it when data points form a swarm is a total waste of time, but you are free to make that "study".

P.S. OP did a linear regression on 3 data points and made some conclusions. That should be enough even for a scientifically aware high school kid to realize what kind of pointless and clueless discussion this whole thread is.

You seem to have a strong emotional reaction to the idea of doing a linear regression between geekbench and SF NPS on various CPUs/SOCs. As I've said before, I would be supportive of such an effort. I even suspect that there's already enough information available to do this task without having to load either SF of Geekbench onto any computers: certainly Geekbench scores are available for a large number of computing devices.

To answer your specific points:

* If the data points form a "swarm", which would result in a low correlation coefficient, it would not be a "total waste of time" - we would then know that Geekbench scores are not a reliable indicator for SF NPS. I suspect that Geekbench is a reliable (or at least "reasonable" indicator). The thing that would dissuade me from this POV would be data, not rhetoric.

* Yes - the OP did do a linear regression on a small sample size. No - the OP did not draw any conclusions from this linear regression result, except to imply that it would be a useful line of enquiry.

btw: one thing I will say with confidence: CPUs which are especially slow on NPS will also get low Geekbench scores, and CPUs which are especially fast on NPS will get high Geekbench scores. This alone almost guarantees a "good" correlation overall. So IMO I am safe in saying that Geekbench numbers will provide a "reasonable" guide to how good a CPU will be at chess.

[Ras]

A laptop running i7-11800h i just purchased
Another laptop I just purchased for my daughter (i5-1135g7)

Laptops are meaningless because what you benchmark isn't the CPU, it's the cTPD configuration of that specific laptop model, and also its cooling solution. In case of AMD laptops, also whether they run on AC or on battery, and for Intel laptops on battery, what the battery charge level is. Also, Apple's M1 is totally out of the equation because Geekbench doesn't use the instructions that make the M1 slow with Stockfish. Basically, that leaves x86 desktop CPUs as only potential comparison targets.

That said, here's the data for my 3400G (non-OC):
Stockfish Ipman bench: 12793 kNPS
Geekbench (multi): 4259

[towforce]

How about... a web page on which you could select the hardware/configuration, which would then guess the NPS number?

Wouldn't be too arduous a task - and once built, would be easy to improve.

If that's too much, it could even be a simple spreadsheet!

[Ras]

How about... a web page on which you could select the hardware/configuration

How about the webpage where you just look up the CPU? See here: http://ipmanchess.yolasite.com/amd--int ... ckfish.php

[towforce]

How about... a web page on which you could select the hardware/configuration

How about the webpage where you just look up the CPU? See here: http://ipmanchess.yolasite.com/amd--int ... ckfish.php

Well that's Xmas sorted then: if anyone asks what to get me, I will tell them I need a smart watch with:

* 512 MPI (message passing interface)

* each of the above having 128 threads

* this giving a total of 512*128=65536 threads

This will enable me to run SF at 49 billion nodes per second, and finally give me a challenging game of chess!

Oh - and the watch should also have all health and fitness tracking functionality, and a week's battery life between charges!

[Milos]

A study, seriously???
There are 3 points in this thread, there are a couple of points for M1 CPU, anyone can generate numbers for their own CPU they have. So your "study" is to collect those data do a linear regression and get a correlation coefficient.
Ofc doing it when data points form a swarm is a total waste of time, but you are free to make that "study".

P.S. OP did a linear regression on 3 data points and made some conclusions. That should be enough even for a scientifically aware high school kid to realize what kind of pointless and clueless discussion this whole thread is.

You seem to have a strong emotional reaction to the idea of doing a linear regression between geekbench and SF NPS on various CPUs/SOCs. As I've said before, I would be supportive of such an effort. I even suspect that there's already enough information available to do this task without having to load either SF of Geekbench onto any computers: certainly Geekbench scores are available for a large number of computing devices.

To answer your specific points:

* If the data points form a "swarm", which would result in a low correlation coefficient, it would not be a "total waste of time" - we would then know that Geekbench scores are not a reliable indicator for SF NPS. I suspect that Geekbench is a reliable (or at least "reasonable" indicator). The thing that would dissuade me from this POV would be data, not rhetoric.

* Yes - the OP did do a linear regression on a small sample size. No - the OP did not draw any conclusions from this linear regression result, except to imply that it would be a useful line of enquiry.

btw: one thing I will say with confidence: CPUs which are especially slow on NPS will also get low Geekbench scores, and CPUs which are especially fast on NPS will get high Geekbench scores. This alone almost guarantees a "good" correlation overall. So IMO I am safe in saying that Geekbench numbers will provide a "reasonable" guide to how good a CPU will be at chess.

Sorry to burst your bubble, but you seem to be pretty clueless what "good" correlation is and what reasonable means at all.
Most ppl with some actual knowledge (and no just superficial rants or trolling) on this forum know where all information is (both SF benches and Geekbenches) and can actually do what you are interested in, but no one is interested simply because they already know the answer.
You seem to be interested in reinventing the wheel (and showing that square is the best shape for it) so please, for a change, go ahead and do something (even so trivial as linear regression of widely publicly available data) yourself, and then come back and present us the result.

[Milos]

That said, here's the data for my 3400G (non-OC):
Stockfish Ipman bench: 12793 kNPS
Geekbench (multi): 4259

Is that with the newest bench, i.e. SF 14.1?
That seems incredibly high for an old gen 4 core CPU. I guess you are runing avx2 SF executable (or maybe even just "modern")?
Or it's a classical SF?

[dangi12012]

I propose we do a study in order to find best benchmark (geekbench 5 integer, cinebench, etc) that highly correlates to chess kn/s (stockfish 14 as the standard).

I have 3 computers
A desktop running i7-9700k
A laptop running i7-11800h i just purchased
Another laptop I just purchased for my daughter (i5-1135g7)

I had a 10th gen i7 laptop but gave it away.

I did a regression analysis for the 3 computers (bench vs geekbench integer). I got r^2 = 0.9675

The regression equation is y = 3.69x -8346 where x is geekbench integer score and y is the bench (command line "bench 1024 xx 26" where xx is the number of threads).

Here are my numbers
i7-9700k 16509kn/s bench and 6317kn/s geekbench integer
i5-1135g7 6934kn/s bench and 4351kn/s geekbench integer
i7-11800h 21538kn/s bench and 8306kn/s geekbench integer

I haven't did a regression on cinebench yet.

Feel free to add your numbers here to get a more accurate regression analysis.

I feel bad for you. The only people who commented are really bad trolls. A forum is to discuss and not only publish results like other people suggested.

Of course you can map from integer iops to chess iops because SF uses IMUL and ADD and Shift with 64 variables mostly. But keep in mind that the new version also uses a lot of AVX2. But I think 90%+ correlations seems correct.

What exact SF version did you use? (Pls link it)
What exact SF command do you run? (Pls add the sf command here? Is it just "bench"?)

Then I will provide my numbers fo a Ryzen 9 5950X.

[towforce]

The only people who commented are really bad trolls.

That's wrong: some of the posts in this thread are supportive, and some contain helpful information.

[Ras]

Is that with the newest bench, i.e. SF 14.1?

Ah, I picked the wrong record referring to asmFishL_2017-05-22_popcnt, that was with the old Ipman bench. With SF 14.1 pop nn-1340 Linux as on Ipman's current page, it's 5969 kNPS.