Yes, it's true that Stockfish on a single core at 5 seconds would not play 45...Qh8.

Stockfish 8 x64

<Nerwal: <If 54...Rd4, then?>

<<The prettiest is 54... ♖d4 55. ♖xf7 ♖xf4 56. ♖g7+ double check and mate.>>>

click for larger view

click for larger view

click for larger view

I have tried to reproduce this on my own computer. Google ran it on 64 threads, yielding like 70 million nps. On 6 cores I get around 8 million nps, so in 9 minutes Stockfish should come up with the same evaluation as it did in 1 minute for Google.

You mean that the real Stockfish gives 49... Rf8 as the move which quickly loses (+6.5) and 49... Kf8= after a few secs.

This could be due to 1 GB of hash tables with 64 threads -- a bad idea, as each thread demolishes the hash table of the other threads, making the extra cores far less useful, and introducing far more chance of error:

Hash tables are how threads speak to each other, but too many using the same memory introduces errors -- a known but ignored (due to its slowdown) situation by chess engines since it is extremely unlikely to occur (and such errors normally have little impact).

I don't know about you, but I found the reference to <threads> strange. To express the degree of parallelism available to a chess engine one typically (always?) refers to the number of <cores> available to the engine. References to 64 <threads> could mean many things, among them:

1. 64 threads executing in a system with 64 cores. The efficiency of the system then depends on whether the cores are tightly or loosely coupled.

2. 64 threads executing in an x86 system with 32 cores capable of hyperthreading with 2 threads executing in each of the 32 cores.

3. A multi-core system with << 64 cores. You could, in theory, run 64 threads in a single-core computer although there is no point in doing so since at any instant in time only one thread could be executing. And the chess engine will take longer to execute than when a single thread is specified since there will be a constant overhead when switching between the various threads.

Other configurations are possible. Without knowing the hardware configuration that Stockfish ran in this "exhibition", it's hard to reach any conclusions.

And remember that multi-core chess engines are notoriously non-deterministic. The same engine, running on the same computer, analyzing the same position, at the same search depth will come up with different move evaluations and possibly different move rankings in subsequent analyses. Not <may>, <will. Guaranteed. So it's not unreasonable that your local Stockfish came up with different suggested best moves than the Stockfish consideration used in the "exhibition".

But I don't understand your reference to one thread demolishing the hash table of the other threads. For best performance it's best that all threads share the same hash table since most of the hash table-related operations are read operations and do not affect the hash table. It's only when new elements are deleted from or added to the hash table that the hash table must be protected from inconsistent modifications and hash pointer disruptions.

How weak is engine's mind!"

Where in London?

It is in fact sited at Google HQ, which is near Tottenham Court Rd Tube station.

Go out of exit three of the Tube, walk over to St Giles' Church👀⛪. Next to the church is The Angel pub🍻, and over the road from that is St Giles Piazza, and at the other end of the piazza is Google HQ.

Go in the front door and take the lift to the ninth floor. On the left-hand side is the <Research and Development> room. Go in there and down to the end and look for <Room 919A>.

AlphaZero (Computer) is in there on the left-hand side. There will be either a man or a woman in there, operating Leela (that is the nickname for AlphaZero (Computer) the men are called either Tom or Stuart; the woman is Rachel.

If a game is in progress they won't let you touch the pieces, but you can twiddle her knobs.