Certain chips can work with useful local models, but compatibility is far behind CUDA.