Benchmark metrics
Warp occupancy
In plain English
How many warps reside on an SM relative to its supported maximum, not how busy its arithmetic units are.
Technical definition
Warp occupancy is the fraction of an SM’s supported resident warp capacity occupied by active resident warps under a kernel’s resource allocation.
Engineering details
Registers, shared memory, block size, and architectural limits constrain residency. Resident warps can still wait on memory, barriers, or dependencies. More occupancy may help hide latency, but reducing registers or useful buffering merely to raise occupancy can make a kernel slower.
Why it matters
Occupancy differs from instruction issue efficiency and from sampled device utilization. Neither a high resident-warp count nor a high utilization percentage proves that tensor cores run near their arithmetic peak.
How to read it in InferenceX
The Blackwell port discusses thread counts and register tradeoffs in copy and attention paths. Interpret those choices jointly with bytes in flight and measured kernel latency, rather than maximizing occupancy in isolation.