15-Way OpenCL Comparison With NVIDIA On Linux, ROCm 1.6 For Radeon
The ROCm 1.6 driver stack does appear to lead to much higher OpenCL kernel latency than the NVIDIA Linux driver.
Rodinia's OpenCL particle filter test was crashing with most of the tested Radeon GPUs, but the RX 580 did perform fine.
Meanwhile, the Lulesh benchmark ported to OpenCL by AMD's GPUOpen project was only performing on ROCm.
Similarly, Xsbench ported to OpenCL by GPUOpen was also only working on Radeon hardware.
Their CoMD OpenCL port was at least working on both vendor GPUs, but appears to be hitting a bottleneck somewhere.