Sorry for this naïve question, but I would like to know in which conditions this value really makes a difference.
Is it only the case with very big depth?
It depends a bit on the engine. But with a good hash table the setting should hardly matter at all when the table is large enough to store some 10% of the reported number of nodes in a typical search. For smaller hash tables the search time increases approximately as the inverse 12th root of the hash side. That is, reducing the hash size by a facttor 4096 would double the search time, which again reduces strength by 50-70 Elo in fast games.
hgm wrote: ↑Thu Nov 12, 2020 4:21 pm
It depends a bit on the engine. But with a good hash table the setting should hardly matter at all when the table is large enough to store some 10% of the reported number of nodes in a typical search. For smaller hash tables the search time increases approximately as the inverse 12th root of the hash side. That is, reducing the hash size by a facttor 4096 would double the search time, which again reduces strength by 50-70 Elo in fast games.
Thanks
Does that mean that running games one at a time with max Threads and max hash size is less efficient than running multiple games in parallel ?
Indeed it does. Doubling the number of threads (or pondering) is not as good as doubling thinking time. Doubling hash has only about 1/12 the effect of doubling thinking time. So the most efficient way to use, say, 8 cores is to run 8 games in parallel, each using 1 thread and 1/8 of the free memory for hash, at a TC 8 times slower than you would otherwise have done.
hgm wrote: ↑Sat Nov 14, 2020 10:55 am
Indeed it does. Doubling the number of threads (or pondering) is not as good as doubling thinking time. Doubling hash has only about 1/12 the effect of doubling thinking time. So the most efficient way to use, say, 8 cores is to run 8 games in parallel, each using 1 thread and 1/8 of the free memory for hash, at a TC 8 times slower than you would otherwise have done.