From Dev.to
Certified kernels, guaranteed at install time
Most GPU inference stacks find their kernel settings on your machine: the first time a new shape shows up, they stop, try block sizes, warp counts and pipeline stages, time them and keep the fastest....