NVIDIA Linux OpenCL Performance vs. Radeon ROCm / AMDGPU-PRO

Starting off with SHOC's Triad test, regardless of the Radeon Linux OpenCL driver used, they came up short of all tested NVIDIA GPUs. With this triad test, the RX 480 and R9 Fury with both drivers were coming up short of a GeForce GTX 1050. Note with the pts/shoc-1.1 test profile used it even had increased the size of the problem/workload, at the request of AMD attributing their slower OpenCL performance to SHOC having too small of a workload.

With the FFT single-precision workload, AMDGPU-PRO does well with the R9 Fury and actually leads it, with using AMDGPU-PRO though and not the "developer preview" of ROCm OpenCL.

And the max single-precision FLOPS.

The AMDGPU-PRO driver was slightly faster than the NVIDIA Linux OpenCL driver for bus speed downloads/read-backs.

But the Radeon OpenCL Linux code really struggled with the texture read bandwidth test where all the results were well short of even a GTX 1050.