I’m inquiring about an application that features two kernels with CUDA synchronization between their executions. Does the cache get flushed after the completion of the first kernel and before the start of the second one?
Kernels launched from the same application do not flush the caches. You can verify this with nsight compute, although nsight compute by default will flush the caches, for profiling purposes, however you can disable that default behavior.
1 Like