HardwareSM
Streaming multiprocessor
In plain English
An NVIDIA GPU execution unit that schedules warps and provides local compute and storage resources.
Technical definition
A streaming multiprocessor is an NVIDIA GPU execution grouping with warp scheduling, registers, shared memory, and arithmetic resources whose details vary by architecture.
Engineering details
A kernel’s resident blocks and warps compete for these resources. Device-wide peak throughput combines many SMs, while instruction measurements may report only one. Available SM count and resource limits constrain how many blocks can run concurrently, but do not determine useful issue efficiency on their own.
Why it matters
Per-SM bandwidth or FLOPs must not be compared directly with a whole-device number. Occupancy, dependencies, instruction shapes, and enabled hardware all affect the fraction of nominal capacity a kernel uses.
How to read it in InferenceX
The Blackwell port reports both per-SM instruction limits and device-level transfer results. Keep the denominator and clock basis explicit when using its figures.