GPU benchmarks can report misleading times when inputs remain in cache between runs. Cache flushing deliberately evicts cached data so each measurement starts from a cold state and reflects the kernel’s memory traffic rather than data retained from earlier executions. section 2.2.10
A practical approach is to zero a buffer larger than the GPU’s L2 cache before each iteration. The buffer must exceed the cache capacity to evict it completely; the benchmark then measures repeated kernel executions under consistent cold-cache conditions. section 2.2.10
