Hashcat has long been a user of OpenCL. I wondered whether that was because CUDA really wasn't much better for this application, but this release puts that to rest:

> One of the biggest advantages of CUDA compared to OpenCL is the full use of shared memory ... This and other optimizations are the reason we improved the performance of bcrypt by 46.90%.

Also interesting that they specifically call out CUDA on ARM devices like the Jetson Nano and Xavier. I suspect that the GPU in the Nano is better than the MX150 in my laptop.