I’d be surprised if it were fruitful. GPUs are basically a few dozen relatively slow cores, with very wide SIMD widths (and also some amount of fixed fn hardwar

I’d be surprised if it were fruitful. GPUs are basically a few dozen relatively slow cores, with very wide SIMD widths (and also some amount of fixed fn hardware, but that’s not super important here). This isn’t at all what you want for any GC algorithms that I know of. That’s ignoring the overhead for copying data between CPU and GPU which I’m unsure you can completely remove even if they share memory.