AMDGPU-PRO 17.10 OpenCL vs. NVIDIA Shows Problems
Last week I began posting a number of AMD Radeon RX 580 Linux benchmarks but not covered so far has been the OpenCL compute performance considering the Clover-based compute stack isn't good enough for benchmarking and is basically unmaintained these days by AMD. Meanwhile, their ROCm stack is still being brought up and is not yet fully-opened nor optimized yet for performance. Thus for those with desktop cards looking for basic OpenCL support are left with the AMDGPU-PRO hybrid driver with its closed-source OpenCL driver. In this article are some fresh OpenCL benchmarks of AMDGPU-PRO on the RX 580 and other Radeon GPUs compared to NVIDIA with its Linux OpenCL driver.
For our latest Linux OpenCL GPGPU comparison are tests on the AMD Radeon side of the RX 480, RX 590, and R9 Fury. On the NVIDIA side is the GeForce GTX 970, GTX 980, GTX 980 Ti, GTX 1050, GTX 1060, GTX 1070, GTX 1080, and GTX 1080 Ti. Tests on the AMD side was done with the latest AMDGPU-PRO 17.10 driver release that brought the RX 500 series support. On the NVIDIA side is the 381.09 Linux beta. All tests were done with the same Intel Core i7 7700K + MSI Z270-A PRO + 16GB DDR4 + Samsung 950 PRO 256GB NVMe SSD. The system was running Ubuntu 17.04 x86_64; with AMDGPU-PRO 17.10 the kernel was downgraded to Linux 4.8 for compatibility with the DKMS driver.
All of these OpenCL benchmarks were conducted in a fully-automated and reproducible manner using the Phoronix Test Suite benchmarking software. But before getting too excited, AMDGPU-PRO 17.10 clearly has an OpenCL regression present, but I went ahead with these results anyhow largely for those curious about the relative NVIDIA Linux GPU performance. In some OpenCL tests, the performance was correct while in others less so:
Triad starts things off a bit slow for AMD relative to NVIDIA with these latest binary Linux drivers...
SHOC's FFT SP test seems to illustrate some AMDGPU-PRO troubles given the incredibly low performance out of the Radeon hardware. This was one of the larger SHOC tests and thus rather surprised to see these numbers so low, clearly some driver issue taking place with the 17.10 series as in the past we haven't seen such abnormal performance with FFT SP.
Yep, it looks like AMDGPU-PRO 17.10 has some OpenCL regressions.
But in a simple test like the bus speed readback with AMDGPU-PRO 17.10, the AMD GPUs were reporting much faster numbers than the tested NVIDIA hardware, too good to be accurate.