I'd like to see hardware more suited for running sparse ML models. You don't need branch prediction like on a CPU but you also don't need as much shared memory as a GPU.