Cloud server vs. HIARCS Pro -

[e4-homie]

This is interesting. According to this video: https://www.youtube.com/watch?v=0GueZpYt7Oo

it takes 20 seconds for a cloud server running Stockfish (700 MN/s) to find the right move (Qxe5). But my Macbook M2 took the same amount of time.

And then the narrator concludes: "This position proves once again that you need Stockfish to run on a more powerful cloud servers".

Am I missing something?

[GoneBallistics]

This is interesting. According to this video: https://www.youtube.com/watch?v=0GueZpYt7Oo

it takes 20 seconds for a cloud server running Stockfish (700 MN/s) to find the right move (Qxe5). But my Macbook M2 took the same amount of time.

And then the narrator concludes: "This position proves once again that you need Stockfish to run on a more powerful cloud servers".

Am I missing something?

Just tried this on my Ryzen 7 5700G, 8 threads, Stockfish 16, 128 MB hash.
Less than 1 second. Position never tried before.
Deep Hiarcs 15.2 4 threads stopped after 30 minutes still Qa3.
How's this possible?

[e4-homie]

I tried again. Fresh restart, HIARCS the only thing running, cleared analysis cache in settings

- With engine DEEP HIARCS. Took less than 2 seconds to find the right move.

- With Stockfish. cleared analysis cache in settings. Took less than 3 seconds to find the right move.

[e4-homie]

Just tried this on my Ryzen 7 5700G, 8 threads, Stockfish 16, 128 MB hash.
Less than 1 second. Position never tried before.
Deep Hiarcs 15.2 4 threads stopped after 30 minutes still Qa3.
How's this possible?

[GoneBallistics]

Just tried this on my Ryzen 7 5700G, 8 threads, Stockfish 16, 128 MB hash.
Less than 1 second. Position never tried before.
Deep Hiarcs 15.2 4 threads stopped after 30 minutes still Qa3.
How's this possible?

I'm stunned too.

[Yarc]

I have in the past found similar non-consistency with chess engines and finding best moves. This was to do with using multiple cores. I setup a chess problem and timed how long each engine took. What I found was that re-testing the same position several times with each engine but clearing the cache each time, the times to solving could vary dramatically from one run to the next. Sometimes the solution would not even be found and other times the solution could be discovered within a few seconds. I think it depends on how the operating system allocates memory for each core and as this can vary from one run to the next accounts for the differences. The only way to get consistent results when testing positions is to use single core. That being said, multi-core engines will always win more points in matches against their single core counterparts.

[GoneBallistics]

I think it depends on how the operating system allocates memory for each core and as this can vary from one run to the next accounts for the differences.

There is another possibility: when you have multiple cores you divide the work between them.
I'm no expert of chess engine internals, but this is what came to my mind thinking about parallele execution.
With many threads (I tried with 8 and 16) this may lead to the best move picked up almost immediately by one of the threads, so it becomes an early candidate and then no move beats it.
In other words, the early picking with several threads may be sort of random (basically what you said), and the probability increases with the number of threads, meaning that the time after which the best move is picked doesn't make too much sense.

[Yarc]

Yes, the work is divided between the cores (including hyper threads that are not as performant as real cores) and the operating system allocates memory to each core/thread as it sees fit, so depending on this memory allocation a particular thread/core that is approaching the correct move may not find it due to less memory available. In this scenario other cores/threads may have more memory and search deeper but along possibly inferior lines. This is one scenario, in another case a particular core/thread may have more memory and rapidly solve the position. A few years ago I did numerous testing on this and my findings were printed in the Selective Search magazine published by Eric Hallsworth. He was interested and asked Mark Uniack about this who said that the only way to get consistent results for puzzle/position solving is by using a single thread. But I also wondered if there is a random-ish affect for a position where more than one move has the same evaluation, if that is possible, in which case how does the engine choose a move? Does it pick one at random? I guess with enough time, it may try one move and then go back and try others. It all depends on the programmer, available time on the clock and memory.

[GoneBallistics]

When you start analyzing a given position and assign the work to several threads you can't do that based on positions you don't know yet.
Let's say you have 16 threads (8 cores, 8 hyperthreads) and 32 possible moves.
Roughly you assign 2 moves per thread (probably simplistic, but makes sense to me).
There's a move clearly much better than all the others, and not so obvious, like in your example.
Is it analyzed as first or as seconds by its thread? Will it be assigned to a real core or to a hyperthread?
When you use a single core, will it be analyzed sooner or later?
You know a lot more than me, but I suspect memory is not the most important factor.
Analyzing with a single core is important trying to determine which engine works better, but chess players want the best move, and they want them ASAP, so they'll throw in the battle all the cores they can afford.

[herO]

This is interesting. According to this video: https://www.youtube.com/watch?v=0GueZpYt7Oo

it takes 20 seconds for a cloud server running Stockfish (700 MN/s) to find the right move (Qxe5). But my Macbook M2 took the same amount of time.

And then the narrator concludes: "This position proves once again that you need Stockfish to run on a more powerful cloud servers".

Am I missing something?

Deep Hiarcs 15.2 didn't solve it even after 4 hours on my computer (32 cores).

[e4-homie]

How is that even possible???

Your CPU is significantly better than my Macbook M2

[e4-homie]

Super curious still about the wide discrepancies reported here. Has anyone else used HIARCS for the diagram in post 1?