Sorts like these work really well on GPUs. You would start with a single thread for the first pass and then keep spawning one-thread-per-bucket for every addit

Sorts like these work really well on GPUs. You would start with a single thread for the first pass and then keep spawning one-thread-per-bucket for every additional pass until done. It isn't particularly taxing on the GPU; your advantage comes from being able to run potentially thousands of threads simultaneously without having to worry about coordinating read/write memory access.