![intel opencl driver performance intel opencl driver performance](https://www.geeks3d.com/public/jegx/2018q4/gpucapsviewer-opencl-support-amd-radeon-intel-gpu-01.png)
These numbers also correlate with 400 and 800 speed differences. Also in 400 case m01000_s04 kernel execution count was around half of the 800 case.
![intel opencl driver performance intel opencl driver performance](https://www.ivanov.biz/wp-content/uploads/2019/04/Screenshot_20190427_164052-cropped.png)
Second one when using export POCL_TRACING=cq the slowest kernel was m01000_s04 in both 400 and 800 MH/s situations but the average runtime in 400 case was around 16 ms and in 800 case around 9 ms. First one is that this speed difference only happens with NTLM hash mode (-m1000). So it was kernel caching issue but I've no idea what caused half of the speed kernel before removing the cache.įew more important details that I remember during testing. Now all versions that I tried earlier were reporting around 800 MH/s consistently. After deleting the cache I wasn't able to reproduce the lower speed anymore by trying out different versions of pocl and LLVM that gave me half of the speed earlier. When trying to figure this out I learned that hashcat actually has a kernel cache under home folder. Then I was able to alternate between 400 and 800 MH/s speed by compiling pocl just by chancing minor version number. Hashcat reported around 400 MH/s with pocl and Intel around 800 MH/s consistently. One interesting thing that I stumbled upon during testing was that at some point I noticed that indeed with pocl got a half of the speed compared to the Intel. I can't try POCL on my own side yet due to a bug. So I'm surprised than POCL is half the speed of Intel, but equivalent to AMD since my experience show Intel and AMD are close, I was naively expecting POCL being half the speed of AMD too if it's half the speed of Intel. I haven't tried oclHashcat but with another benchmark (luxmark), the AMD and Intel results are close on the same Intel CPU. If POCL and AMD results are pretty close on an AMD CPU, will they be close together on an Intel CPU too? Is Intel twice faster than AMD too on their Intel CPU, or is POCL half slower than the two others? One interesting experiment would be to try the AMD OpenCL stack with your Intel CPU, so you will get three results for the same CPU, using the Intel stack, using the AMD stack, using the POCL stack. The speed between the native AMD CPU OpenCL runtime (part of Catalyst) and pocl is pretty close to each other that's why the results for Intel is very interessting. I did the same test on my AMD FX-8120 CPU.