That’s most likely an issue of drivers. But having a pure CPU path in addition of the GPU one is the main performance bottleneck right now in mainstream dt. Just open darktable -d opencl -d perf and look at how much time is lost in the control queue and in the buffer copies… it’s between half and 3/4 of the pipeline runtime depending on the modules you use. That’s completely insane for something that will be thrown to the screen by the GPU in any case at the end.