CUDA Programming

Occupancy

Measure of the ratio between the number of active warps and the maximum number of warps that can be resident on a Streaming Multiprocessor. High occupancy does not necessarily guarantee better performance but helps hide latency.

← Geri