Yeah this is the trick. You need to maximize the use of workgroup parallelism and also lay things out in memory for those kernels to access efficiently. It’s a bit of a balancing act and I’ll be working on benchmarks to test out different strategies.