The problem is that CUDA already does that, and they're not incentivized to really improve from that baseline, given the capability and ease of use of the ROCM or Intel solutions.