Cuda optimization actually doesn’t suck that much. I think NSight studio is amazing and super helpful for profiling and identifying bottlenecks in kernels