I'm a Python numba.cuda user, I think this is the end of my exploration. It's quite disappointing that Apple doesn't care about non-apple researcher at all. PyTorch and Tensorflow are two high-level language for any specific calculations, while C++/OC-based GPU coding is too much for non-CS engineering student. Swift-based metal shader language would be a brilliant idea. I will keep using my 3090 for GPU calculations and leave the M1 Max for lighting up the displays only.