If you’re training a model that needs 128gb, the GPU speed is going to be a “bottleneck”

I believe that Apple Silicone is still too slow for training. The big RAM is really useful for interference of big models though