afw has GPU implementations of convolution and warping that are optionally built with nvidia's CUDA runtime. I don't believe they're regularly built by anyone, and they're certainly not included in official CI, and that makes it likely they've bitrotted. I also think it's unlikely that we'll add more GPU implementations for other LSST code in the future, making these not-very-useful even to someone who did want to run with GPUs.
I propose we remove them and all of the code that deals with conditional CUDA builds and configuration. If we do want to find them in the future, we'll always have our git history.