FARSH: hashing 30 GB/s!

[Bulat Ziganshin]

JamesB, i have added anti-optimization trick to loop measuring speed for the first column. Please try it with compiler/options that produced strange results

Overall, it seems that newer Intel CPUs from IvyBridge to Skylake has pretty close speed/frequency.

We still miss Nehalem, SandyBridge, low-power CPUs (two generations of Atoms, Bobcat, Jaguar), Pentium 4, original Athlon64 and any older CPUs

[Piglet]

AMD Sempron 3000+ Socket A (462) 2000 Mhz (real clock 1988 Mhz) http://www.cpu-world.com/CPUs/K7/AMD...A3000BOX).html

C:\teszt\farsh02-benchmark>for %e in (*.exe) do @start /b /wait  /realtime %e 1
..
aligned-farsh-x86         |   2.057 GB/s =  1.916 GiB/s  |   2.536 GB/s =  2.362 GiB/s
AMD Sempron(TM) 3000+: SSE
..
farsh-x86                 |   2.022 GB/s =  1.884 GiB/s  |   2.348 GB/s =  2.187 GiB/s

[hexagone]

For Sandy Bridge:

for %e in (*.exe) do @start /b /wait /realtime %e 1
aligned-farsh-x64-nosimd | 12.477 GB/s = 11.620 GiB/s | 14.035 GB/s = 13.071 GiB/s
aligned-farsh-x64 | 19.638 GB/s = 18.289 GiB/s | 25.462 GB/s = 23.714 GiB/s
aligned-farsh-x86-sse2 | 22.342 GB/s = 20.808 GiB/s | 26.134 GB/s = 24.339 GiB/s
aligned-farsh-x86 | 5.088 GB/s = 4.738 GiB/s | 5.834 GB/s = 5.433 GiB/s


Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz: AVX
farsh-x64-nosimd | 12.061 GB/s = 11.233 GiB/s | 13.780 GB/s = 12.834 GiB/s
farsh-x64 | 21.969 GB/s = 20.460 GiB/s | 25.183 GB/s = 23.454 GiB/s
farsh-x86-sse2 | 20.794 GB/s = 19.366 GiB/s | 24.705 GB/s = 23.008 GiB/s
farsh-x86 | 5.098 GB/s = 4.748 GiB/s | 6.012 GB/s = 5.599 GiB/s

[JamesB]

A very quick test of FARSH on a AVX-512 system. There's no dedicated AVX-512 version, so I tried the normal avx2 and also non-simd variant with arch=core2 vs arch=native. Odd how it's so much slower with icc.

gcc 6.2.0:
farsh-x64-avx2            |  10.529 GB/s =  9.805 GiB/s  |  12.352 GB/s = 11.504 GiB/s
farsh-x64<core2>          |   2.514 GB/s =  2.341 GiB/s  |   2.416 GB/s =  2.250 GiB/s
farsh-x64-<native>        |   7.415 GB/s =  6.905 GiB/s  |  29.285 GB/s = 27.274 GiB/s

icc 16.0.2:
farsh-x64-avx2            |   6.712 GB/s =  6.251 GiB/s  |   7.776 GB/s =  7.242 GiB/s
farsh-x64-<native>        |   3.998 GB/s =  3.724 GiB/s  |   4.262 GB/s =  3.969 GiB/s



model name    : Intel(R) Xeon Phi(TM) CPU 7210 @ 1.30GHz

[Bulat Ziganshin]

seems gcc successfully optimized main cycle for avx-512? with 72 cores, overall speed should be 2.1 TB/s

[JamesB]

It wasn't multi-threaded

I'm not sure at what speed the memory can be driven, but I believe there is distributed memory per core so in theory you can parallelise that too. Getting data into the per-core memory would then be the bottleneck. I think we're over-doing it a bit though.

Anyway, this machine "only" had 64 cores (x4 hyperthreads).

[Razor12911]

Intel(R) Core(TM)2 Extreme CPU X9650 @ 3.00GHz: SSE 4.1

aligned-farsh-x64-nosimd  |   9.655 GB/s =  8.992 GiB/s  |   7.473 GB/s =  6.960 GiB/s
aligned-farsh-x64         |  15.810 GB/s = 14.725 GiB/s  |  19.617 GB/s = 18.269 GiB/s
aligned-farsh-x86-sse2    |  17.291 GB/s = 16.104 GiB/s  |  19.806 GB/s = 18.446 GiB/s
aligned-farsh-x86         |   4.640 GB/s =  4.321 GiB/s  |   5.449 GB/s =  5.075 GiB/s
farsh-x64-nosimd          |   5.314 GB/s =  4.949 GiB/s  |   5.551 GB/s =  5.170 GiB/s
farsh-x64                 |   6.789 GB/s =  6.322 GiB/s  |   7.865 GB/s =  7.325 GiB/s
farsh-x86-sse2            |   6.872 GB/s =  6.400 GiB/s  |   8.898 GB/s =  8.287 GiB/s
farsh-x86                 |   3.930 GB/s =  3.660 GiB/s  |   4.725 GB/s =  4.401 GiB/s

Intel(R) Celeron(R) CPU 1005M @ 1.90GHz: SSE 4.2

aligned-farsh-x64-nosimd  |   8.419 GB/s =  7.841 GiB/s  |   8.984 GB/s =  8.367 GiB/s
aligned-farsh-x64         |  13.252 GB/s = 12.342 GiB/s  |  16.139 GB/s = 15.031 GiB/s
aligned-farsh-x86-sse2    |  12.614 GB/s = 11.747 GiB/s  |  15.991 GB/s = 14.893 GiB/s
aligned-farsh-x86         |   2.941 GB/s =  2.739 GiB/s  |   3.318 GB/s =  3.090 GiB/s
farsh-x64-nosimd          |   8.207 GB/s =  7.644 GiB/s  |   8.985 GB/s =  8.368 GiB/s
farsh-x64                 |  13.514 GB/s = 12.586 GiB/s  |  15.595 GB/s = 14.524 GiB/s
farsh-x86-sse2            |  12.100 GB/s = 11.269 GiB/s  |  15.822 GB/s = 14.736 GiB/s
farsh-x86                 |   2.894 GB/s =  2.695 GiB/s  |   3.467 GB/s =  3.229 GiB/s

Intel(R) Atom(TM) CPU Z3735G @ 1.33GHz: AES-NI

aligned-farsh-x86-sse2    |   2.713 GB/s =  2.526 GiB/s  |   3.260 GB/s =  3.036 GiB/s
aligned-farsh-x86         |   1.336 GB/s =  1.244 GiB/s  |   1.621 GB/s =  1.509 GiB/s
farsh-x86-sse2            |   3.178 GB/s =  2.960 GiB/s  |   3.734 GB/s =  3.478 GiB/s
farsh-x86                 |   1.384 GB/s =  1.289 GiB/s  |   1.632 GB/s =  1.520 GiB/s