Processor Benchmarks

[Rebel]

https://www.cpubenchmark.net/compare/In ... 3099vs2905

Makes a comparison between 3 CPU's.

Single Thread Rating
Intel i7-860 = 1229
Intel i7-7700 = 2350
Intel i7-8700 = 2630

To what extent can you trust that for chess the i7-8700 will run 2.13 x faster than the i7-860?

2630 / 1229 = 2.13

[Joost Buijs]

It depends solely on what you are going to do with it. I assume they are taking multithreaded performance into account (6 cores vs 4 cores), for a single threaded program the difference will be less. The ICP improved maybe 20% over the last 8 years, so the difference in clock frequency + 20% will be a good estimate for a single threaded program.

[Jesse Gersenson]

I recently had this exact conversation with Jiri Dufek, using the same cpubenchmark.net site, and determined the benchmarks aren't useful for determining chess performance.

But, let's get real numbers. Anyone here have any of the above cpu's? If so, run the stockfish 9 benchmark and post your results.

[Werewolf]

It's really easy to work this out, especially if HT is off.

We know the core count, N.

We know the all-core turbo clock speed, CS.

We know the processor performance co-efficient. (basically Nehalem = Westmere = Sandy Bridge = Ivy Bridge, so let that be 1.0 Then everything from Haswell onwards is about +14% IIRC)

So something like:

CS x (N cores ^0.8) x co.ef

Then you have an absolute speed reference and we know a doubling of speed is x elo on today's engines (60? I forget)

[Rebel]

It depends solely on what you are going to do with it. I assume they are taking multithreaded performance into account (6 cores vs 4 cores), for a single threaded program the difference will be less. The ICP improved maybe 20% over the last 8 years, so the difference in clock frequency + 20% will be a good estimate for a single threaded program.

Suspected something like that already.

Made a script.

http://rebel13.nl/home/cbl.html

Perhaps some of you can contribute or comment if it makes sense to do so.

[Joost Buijs]

About your script:

The Stockfish bench produces a lot of redundant console output which can have a big influence on the performance shown on certain systems.
To get a more realistic result it is better to dump the bulk output to /dev/null, e.g. open a command window and type 'stockfish9.exe bench 128 1 20 > NUL'. The final result is sent to stderr so that will still show up in the console window.

[Rebel]

Looks better indeed.

Made it more easier. Also because I could derive from the entries the page did not excel in clearness.

http://rebel13.nl/home/cbl.html

I will leave it as is for the moment, maybe tomorrow make it official with a clean start.

Thanks all for feedback!

[Guenther]

Looks better indeed.

Made it more easier. Also because I could derive from the entries the page did not excel in clearness.

http://rebel13.nl/home/cbl.html

I will leave it as is for the moment, maybe tomorrow make it official with a clean start.

Thanks all for feedback!

http://adamsccpages.blogspot.com/p/stoc ... hmark.html
(done with SF8 bench)

[Rebel]

Interesting... since the unpredictable nature of multithreading maybe it's indeed even better to ask for the NPS instead of the time elapsed.

The bench.zip file is no longer available, I assume it states how to run the SF8 benching, do you still have it?

I will link to Adam's page anyway.

[Guenther]

Interesting... since the unpredictable nature of multithreading maybe it's indeed even better to ask for the NPS instead of the time elapsed.

The bench.zip file is no longer available, I assume it states how to run the SF8 benching, do you still have it?

I will link to Adam's page anyway.

Yes I still have it. I guess we should inform Adam first, that the download doesn't work anymore.
Probably he has not enough time since a while, because the list wasn't updated for long
and I know that even my own results are not added until now. (sent one year ago)

Guenther