The problem is that this is an "inferencing" accelerator - i.e. it can only execute pretrained models. You cannot train a model on one of these, you need a trai

The problem is that this is an "inferencing" accelerator - i.e. it can only execute pretrained models. You cannot train a model on one of these, you need a training accelerator. And pretty much all of those are either NVidia GPUs or cloud-only offerings.