[Computer-go] Xeon Phi result

[Gian-Carlo Pascutto]

Hi all,

I managed to get a benchmark off of a Intel® Xeon Phi™ Processor 7250
16GB, 1.40 GHz, 68 core (272 thread) system.

I used a version of Leela essentially identical to the public Leela
0.10.0, but compiled with -march=knl (using gcc 5.3), using an
appropriate version of Intel MKL (2017.1 for MIC) and increasing the
maximum amount of threads.

benchmark:

~ 151000 g/s (557 g/s per thread)

netbench:

predictions ->   670 p/s
evaluations ->  3007 p/s

This was with the 16G HBM bound as addressable memory. Using the regular
DDR4 cuts "netbench" numbers in half, but has no big impact on "benchmark".

This means it's about 5 times faster (in integer operations) than a quad
core desktop with HT, and about similar in floating point performance to
a mid-range videocard.

It's a nice machine, if a bit pricey.

-- 
GCP