Ditching CUDA would mean accepting at least 10-30% performance hit on NVIDIA cards and losing the ability to use the newest NVIDIA improvements. OpenCL doesn't

Ditching CUDA would mean accepting at least 10-30% performance hit on NVIDIA cards and losing the ability to use the newest NVIDIA improvements. OpenCL doesn't know about HMMA (aka "Tensor Cores"), so it would be about an order of magnitude slower for DL on Volta. I think you'd be hard pressed to find people willing trade significant performance for the option of running applications on AMD's less power efficient processors.