E5-2650 / RX 580 4GB

[r0k0]

finding it quite odd that my RX580 4GB with 1x E5 2650 2.5Ghz boost are pushing 1 500 000 PPD, is this normal ? when i go with my 10700K / 2080ti i'm over 2 millions....

[Neil-B]

The time will be using cuda which gives it a boost compared to amd and your 10700k will be clocking faster than the e5 .. doesn't surprise me as such but haven't checked the averages for the cards .. actually if looks like that might be quite good performance for the rx580

[r0k0]

2hours later it drop to 350K PPD

[JimboPalmer]

There are several CPUs called the E5 2650, v2 only knows AVX while v4 knows AVX2 so is about 40% faster for F@H (AVX can do a ADD or a MULT at one time, while AVX2 can do an ADD and a MULT in one step)

I interpret OpenCL and CUDA differently

AMD is all in on OpenCL it is their prime way of doing general purpose math on their GPUs.
For Nvidia, CUDA is their preferred, tuned way of general purpose math, (it locks you into a Nvidia GPU, which is ideal if you are Nvidia) while OpenCL is just some other standard they met once. (There is no sign Nvidia has updated OpenCL support since 2015) CUDA support continues to be tuned for newer GPUs.

For F@H, OpenCL is the latest best way to use an AMD card but OpenCL on Nvidia does not use any new features on current GPUs that development is just on CUDA.

Some AMD drivers support OpenCL 2.2, most support 2.0, no Nvidia OpenCL driver is beyond OpenCL 1.2.
Nvidia has updated CUDA from 3.5 to 8.6 without updating OpenCL.

[r0k0]

does compute mode or graphic mode change the PPD ?

[bruce]

FAH does not use the graphics hardware; it only uses the compute portions of the GPU.

JimboPalmer: The FAHCore which does the OpenCL calculations for either AMD or nV require 1.2. As far as I know, it doesn't make use of any of the newer features in 2.x except if they happen to enhance the performance of a function that already existed in code written for 1.2.