4 ms·

According to the docs, it would be more like CUDA / OpenCL. I think there are some OpenCL bindings for it.