@dneto Ideally very close to theoretical peak. Basically problems with "interesting" parallelism including 2D rendering of course but also including sparse matrix multiplication, parsers, and so on. Yes, but that's at a much higher level than what I'm talking about; Futhark and Taichi are probably closer to the mark as far as languages that might compile down to run on GPC (as of course would be PyTorch and the MLIR ecosystem).

@aras Right, that's what basically started the exploration. I'm unhappy with them because of three limitations: no joins (yet, though that's planned), fixed size data only, and no ordering guarantees. That makes them pretty much unsuitable for Vello. The centerpiece of my idea is that the launch logic is programmable, so you can express joins and other things.