Ram usage

[Krzysztof Grzelak]

Thanks for the other replies.

[RubiChess]

I may be in the minority here, but I find these huge hash sizes (e.g. 64GB, 128GB) ridiculous. Just because you can does not mean you should, and where is the evidence that it improves engine performance and strength ?

And why do you think so.I understand that for a TCEC tournament they would give a hash of 1024, it would be okay for you.

From the TCEC Wiki: https://wiki.chessdom.org/TCEC_Season_F ... superfinal

"RAM: 256GiB (~96GiB/engine)"

This may also answer your original question.

[Graham Banks]

For 16 cores, I would use 6GB hash tops.

For 32 cores, I'd use 12GB hash tops.

That would apply to long time controls like 2 hours with 30 second increments.

[RubiChess]

For 16 cores, I would use 6GB hash tops.

For 32 cores, I'd use 12GB hash tops.

That would apply to long time controls like 2 hours with 30 second increments.

But you should consider that many engines can only use 2^n entries (buckets) in the hash table which usually (assuming entry/bucket size also some 2^m bytes) means that it would only use 4GB when you select 6GB hash size and 8GB when you select 12GB.

SF is one of the few engines that don't need a power-of-2 hash size.

Regards, Andreas

[Graham Banks]

For 16 cores, I would use 6GB hash tops.

For 32 cores, I'd use 12GB hash tops.

That would apply to long time controls like 2 hours with 30 second increments.

But you should consider that many engines can only use 2^n entries (buckets) in the hash table which usually (assuming entry/bucket size also some 2^m bytes) means that it would only use 4GB when you select 6GB hash size and 8GB when you select 12GB.

SF is one of the few engines that don't need a power-of-2 hash size.

Regards, Andreas

For CCRL time controls, I think that 4GB and 8GB would suffice, but for longer time controls, I'd allocate more to be safe.

[yurikvelo]

That would apply to long time controls like 2 hours with 30 second increments.

This is not initial goal, author asked about engine testing, not about infinite analysis

[Graham Banks]

That would apply to long time controls like 2 hours with 30 second increments.

This is not initial goal, author asked about engine testing, not about infinite analysis

I'm talking about engine v engine testing.