TalkChess.com

For chess engines which is better ? clock speed or core count ?

For example

20 cores @4000 GHz or 40 cores @3000 GHz

Clock speed is more important because the more cores involved in search, the more search inefficiency you introduce.

However, the example you give is actually faster for the 40 core machine because it’s not a fair comparison.
Fairer would be 20 cores at 4000 GHz or 40 cores at 2000 GHz, and in this instance the 20 core machine would win.

In your example you’ve raised the larger machine’s clock speed to 3000 GHz which is enough for it to pull ahead.

It depends on the particular engine as well. Not all of them can take effective advantage of high core counts, so for those cases the lower core count at higher frequency is better.

It also depends on what you are doing with the engines.

For instance, when I use my 64 core machine to analyze, I usually do not use 60+ cores on a single position at a time.
Instead, I break the file to be analyzed into a collection of sub-files and then divide the thread count appropriately.

So, suppose I have 60K positions to analyze.
I might divide them into 6 files of 10K positions each.
Then I start up 6 processes each with 10 threads (maybe one or two with 11) so that I can analyze all of the sub-files at the same time.
This is definitely faster than analyzing the entire original file with all the threads at once.

(I always leave at least one or two cores free so that the machine is still responsive).

Clock speed.

I ran a lot of benchmarks .. This is my conclusion. This is the simple formula for the current engines & systems.
The higher the Clock speed x core the better.

40 x 3000 = 120000
20 x 4000 = 80000

@Werewolf was kind of right.

If this was the case
40 x 2000 = 80000
20 x 4000 = 80000

the the time control comes into play.
STC : 20 x 4000 is better.
LTC : both are the same.

Or you can check Ipman's benchmark list.

http://www.ipmanchess.yolasite.com/amd- ... -bench.php

JP.

lovetb wrote: ↑Sun Jan 20, 2019 8:35 am I ran a lot of benchmarks .. This is my conclusion. This is the simple formula for the current engines & systems.
The higher the Clock speed x core the better.

40 x 3000 = 120000
20 x 4000 = 80000

@Werewolf was kind of right.

If this was the case
40 x 2000 = 80000
20 x 4000 = 80000

the the time control comes into play.
STC : 20 x 4000 is better.
LTC : both are the same.

It's not quite this simple I'm afraid. What this is telling you is how many nodes the engine is looking at. What it's not telling you is how efficient the search is as it looks at these nodes.

The more cores used, the less efficient the search. So if you had one core running really fast looking at 10 million nps and 10 cores running slowly looking at 10 million nps - the single core would win.

However, in the last few years this doesn't seem to be as pronounced as it was in the past.

Werewolf wrote: ↑Sun Jan 20, 2019 9:20 am
However, in the last few years this doesn't seem to be as pronounced as it was in the past.

That is why I said "This is the simple formula for the current engines & systems".

lovetb wrote: ↑Sun Jan 20, 2019 9:40 am

Werewolf wrote: ↑Sun Jan 20, 2019 9:20 am
However, in the last few years this doesn't seem to be as pronounced as it was in the past.

That is why I said "This is the simple formula for the current engines & systems".

OK, but just measuring the nps could be improved on, especially if you're running a dual xeon with lots or cores & threads, e.g with one of the logarithmic attempts.

Anyway, it's your system so it's up to you but getting it right will make a difference when you're deciding between how many nodes to apportion to a position or whether to switch on HT or not.