Ipman has 2 interesting SF benchmarks

[Jouni]

102.542.506.414 nps for 131.072 threads cluster!
85.835.086 nps for AWS Graviton3 64 threads. What's this?

[Vinvin]

102.542.506.414 nps for 131.072 threads cluster!

https://ipmanchess.yolasite.com/amd--in ... ckfish.php

It would be interesting to have a short video (even took with a smartphone) of this benchmark.
And a short video (15 to 30 seconds) displaying NPS in the starting positions with this monster.

[CornfedForever]

I've been wondering if there is a rule of thumb regarding NPS and # of threads. Ex:

NPS: 69.539.701 AMD Ryzen 9 7950X @4.9Ghz DDR5 6000 CL30 32threads

I doubt it would be fair to say therefore:
16 threads would = about 34,769.850 NPS
and 8 threads = 17,384,925 NPS

But perhaps it is NOT a linear relationship like that. Would anyone have a 'rule of thumb' if not?

[smatovic]

[...]
85.835.086 nps for AWS Graviton3 64 threads. What's this?

Amazon AWS cloud Graviton3 CPU, ARM Neoverse V1 based with 64 cores (up to two sockets?):

https://en.wikipedia.org/wiki/AWS_Graviton#Graviton3

https://en.wikipedia.org/wiki/ARM_Neoverse#Neoverse_V1

--
Srdja

[smatovic]

I've been wondering if there is a rule of thumb regarding NPS and # of threads. Ex:

NPS: 69.539.701 AMD Ryzen 9 7950X @4.9Ghz DDR5 6000 CL30 32threads

I doubt it would be fair to say therefore:
16 threads would = about 34,769.850 NPS
and 8 threads = 17,384,925 NPS

But perhaps it is NOT a linear relationship like that. Would anyone have a 'rule of thumb' if not?

Generally, Stockfish NNUE NPS depends also on the network size, currently it makes 1M to 2M NPS per core on modern CPUs (depends for example on frequency and vector-unit like SSE, AVX2, NEON), and with HyperThreading on (SMT2, 2 threads per core) you gain roughly 1.5x more NPS.

--
Srdja

[Werewolf]

102.542.506.414 nps for 131.072 threads cluster!

Presumably this is a cloud setup he has access to? I'd love to know the elo gain...

[Jouni]

Obviously confidential tester has access to one TOP500 supercomputer and can run SF bench!

[Jouni]

In discord is even faster bench:

graviton3 / ARM64
sf 15:
git checkout e6e324eb28fd49c1fc44b3b65784f85a773ec61c
make -j ARCH=armv8 profile-build
./stockfish bench 128 64 24 default depth
===========================
Total time (ms) : 57573
Nodes searched : 7099323910
Nodes/second : 123309952

Threadripper speed .

[CornfedForever]

I've been wondering if there is a rule of thumb regarding NPS and # of threads. Ex:

NPS: 69.539.701 AMD Ryzen 9 7950X @4.9Ghz DDR5 6000 CL30 32threads

I doubt it would be fair to say therefore:
16 threads would = about 34,769.850 NPS
and 8 threads = 17,384,925 NPS

But perhaps it is NOT a linear relationship like that. Would anyone have a 'rule of thumb' if not?

Generally, Stockfish NNUE NPS depends also on the network size, currently it makes 1M to 2M NPS per core on modern CPUs (depends for example on frequency and vector-unit like SSE, AVX2, NEON), and with HyperThreading on (SMT2, 2 threads per core) you gain roughly 1.5x more NPS.

--
Srdja

So, in general, in doubling the thread count (say from 8 to 16) instead of a 2x gain in NPS, you get closer to 1.5x gain in NPS - instead of from say 100 to 200, you actually get closer to 150?

[smatovic]

I've been wondering if there is a rule of thumb regarding NPS and # of threads. Ex:

NPS: 69.539.701 AMD Ryzen 9 7950X @4.9Ghz DDR5 6000 CL30 32threads

I doubt it would be fair to say therefore:
16 threads would = about 34,769.850 NPS
and 8 threads = 17,384,925 NPS

But perhaps it is NOT a linear relationship like that. Would anyone have a 'rule of thumb' if not?

Generally, Stockfish NNUE NPS depends also on the network size, currently it makes 1M to 2M NPS per core on modern CPUs (depends for example on frequency and vector-unit like SSE, AVX2, NEON), and with HyperThreading on (SMT2, 2 threads per core) you gain roughly 1.5x more NPS.

--
Srdja

So, in general, in doubling the thread count (say from 8 to 16) instead of a 2x gain in NPS, you get closer to 1.5x gain in NPS - instead of from say 100 to 200, you actually get closer to 150?

If thread and core count is equal you can expcect an doubling of NPS, 8cores with 8threads to 16cores with 16threads doubles the NPS, assuming same architecture and frequency cos most engines scale linear NPS wise nowadays across cores on a single socket, then you turn SMT resp. HT on, and you get further ~1.5x NPS by this, depends, on architecture and engine.

https://en.wikipedia.org/wiki/Simultane ... mentations

Modern Intel and AMD processors profit both from two threads per core, 2-way SMT, in varying percentages, some people prefer SMT resp. HT off during testing.

--
Srdja