brmarkus
commented
1 year ago With or without having a model-cache and/or OpenCL configured - the first run includes compilation of the OpenCL-kernels; depending on how powerful the underlying HW is this can take several seconds or "much longer" (e.g. 30s, 60s or even longer).
When having a cache configured then only the first start is delayed by the OpenCL-kernel-compilation; consecutive runs will used the cached kernels. Otherwise each run will require compilation.
(required for GPU or VPUs, not required for e.g. CPU)
I have an intel Arc A770 and when I use the gpu option for some reason it uses 100% of my cpu and when I use the cpu option it does not use 100% of my cpu.