The next revolution in computer chess?

[smatovic]

Let's not forget the newer CPUs with AVX 512 bit, plus VNNI extensions, plus BF16 extensions, with AMD Zen still on AVX2 256 bit...still some speedup in pipe, or alike.

***edit***

There is also a trend to two vector untis per core, A64FX and some Intel Xeons for example...

--
Srdja

[cma6]

better invest in some AVX based CPU too
Srdja


Srdja:
I see that the NNUE engines are offered in a choice between .bmi , .popcount or .avx
Is there any reason to prefer one of the other?
Thanks, CMA

[smatovic]

better invest in some AVX based CPU too
Srdja


Srdja:
I see that the NNUE engines are offered in a choice between .bmi , .popcount or .avx
Is there any reason to prefer one of the other?
Thanks, CMA

I am not into the concrete SF code and NNUE implementation, but AFAIK the NNUE
Neural Network part profits from being run on SSE or AVX. These are vector units
in the CPU core which can run multiple operations at the same time in SIMD
manner, so the more bit width you have, SSE = 128, AVX2 = 256, AVX-512, the more
throughput you get. GPUs have multiple SIMD/Vector units coupled to Compute
Units, so these are in a way compareable to AVX, beside that AVX is not
connected via PCIe but is part of the CPU and therefore the latencies are lower,
you can call the neural network eval faster, so to speak.

I am not sure what the difference between .bmi, .popcount and .avx SF compiles
are, maybe someone else can clarify which compiles use what for SF AB search
and what for the SF NNUE NN part...

--
Srdja

[Ovyron]

Is there any reason to prefer one of the other?

Yes, some take advantage of new hardware architecture, you'd want to try all the ones that you can run (the others will exit with error, unless you can run them all.) Run the exe without GUI, type bench and hit enter, do it twice and take the average nodes per second. Use the one with the highest average, and that's it.

[cma6]

Thanks, Ovyron, probably that means that I should be looking at AMD TR 3960X or TR 3970X.