10-04-2026, 07:33 AM
Interesting, I hadn't tried that. I don't think your conclusion is far off. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
Mostly CUDA, sometimes wrong, usually measured