Monday, August 26, 2013

Intel CPU only OpenCL benchmark.

I used John and ImageMagick to do the benchmark with the Intel OCL sdk installed.

The results (John) --

mscash2-opencl --
OpenCL platform 0: Intel(R) OpenCL, 1 device(s).
Using device 0:         Intel(R) Core(TM) i3-2120 CPU @ 3.30GHz
Optimal Work Group Size:32
Kernel Execution Speed (Higher is better):0.000444
Benchmarking: M$ Cache Hash 2 (DCC2) PBKDF2-HMAC-SHA-1 [OpenCL]... DONE
Raw:    418 c/s real, 105 c/s virtual

mscash2 --
Benchmarking: M$ Cache Hash 2 (DCC2) PBKDF2-HMAC-SHA-1 [128/128 AVX intrinsics 4x]... (4xOMP) DONE
Raw:    2464 c/s real, 620 c/s virtual

wpapsk-opencl --
OpenCL platform 0: Intel(R) OpenCL, 1 device(s).
Using device 0:         Intel(R) Core(TM) i3-2120 CPU @ 3.30GHz
Benchmarking: WPA-PSK PBKDF2-HMAC-SHA-1 [OpenCL]... (4xOMP) DONE
Raw:    790 c/s real, 199 c/s virtual

wpapsk --
Benchmarking: WPA-PSK PBKDF2-HMAC-SHA-1 [32/64]... (4xOMP) DONE
Raw:    924 c/s real, 232 c/s virtual

bf-opencl
OpenCL platform 0: Intel(R) OpenCL, 1 device(s).
Using device 0:         Intel(R) Core(TM) i3-2120 CPU @ 3.30GHz
****Please see 'opencl_bf_std.h' for device specific optimizations****
Benchmarking: OpenBSD Blowfish (x32) [OpenCL]... DONE
Raw:    1131 c/s real, 381 c/s virtual

bf
Benchmarking: OpenBSD Blowfish (x32) [32/64 X2]... (4xOMP) DONE
Raw:    2448 c/s real, 613 c/s virtual

With ImageMagick convolve filter --
With OpenCL --

convert -convolve -1,0,-1,0,4,0,-1,0,-1 background.png test.png
real    0m5.175s
user    0m5.776s
sys     0m0.064s

Without OpenCL (OpenCL excluded during compile) --

convert -convolve -1,0,-1,0,4,0,-1,0,-1 background.png test.png
real    0m4.985s
user    0m5.560s
sys     0m0.052s


As we can see, the Intel SDK is of no good. Maybe they could've spend more time on developing their hardware and drivers instead of making shiny GUI on .NET.