I have this same issue with my new MacBook Pro 14 with M1 Max fully loaded. I've tried creating a clean python 3.8 and 3.9 installation following instructions here and elsewhere. Tried downgrading my tensor flow-metal package. Just about every possible combination in clean environments and new installations. Bottom line is that while the GPU "works", it runs about 5x slower than running on pure CPU. Aka if I uninstall tensorflow-metal package the same training that took say 11 sec with the package installed takes only about 2.5 sec without the metal package. You can also replicate same results with forcing tensorflow to run on CPU with the metal package installed. Looking at Activity Monitor during run suggests that the M1 Max GPU is in fact loaded with the package installed. It just performs horribly poorly, in fact so badly as to be unusable. My working assumption is that this is not the intended performance, but a bug. What's concerning is that no maintainer in any of the forums, whether it be tensorflow/keras, or whether it be Apple forums, has really acknowledged that this is a bug. Perhaps there's confusion between the different manifestation of the bug in Intel vs. older M1 vs. newer M1 Pro/Max, as well as the different operating systems involved. So let me be unambiguously clear: none of the stuff listed above here, or in other threads, involving reinstallation, downgrading packages, etc. makes this work on my M1 Max properly.