ROCm is much less mature of a deep learning stack than what nvidia has (by atleast 5 years). However your choice depends on what your scope of usage is, and if you’re willing to trade off usability for cost (I’m assuming amd hardware is significantly cheaper).