AI inference glossary

Software

SASS

In plain English

The architecture-specific NVIDIA GPU machine instructions inspected when studying what a compiled kernel actually executes.

Technical definition

SASS names the architecture-specific machine instruction representation used to inspect compiled NVIDIA GPU kernels.

Engineering details

Unlike virtual PTX, it reflects the target instruction stream after compilation. Disassembly can reveal load paths, tensor instructions, and scheduling details absent from high-level source. Instruction names and encodings are architecture-specific, and the presence of an instruction alone does not reveal its achieved throughput.

Why it matters

Generated-code inspection helps explain unexpected performance differences and verify whether a compiler selected the intended path. It complements profiling rather than replacing measurements of dependencies, resource contention, and execution time.

How to read it in InferenceX

The Blackwell article distinguishes LDS from generic LD and discusses LDGSTS, UTMALDG, and UBLKCP. Its conclusions attach to those tested instructions and configurations rather than every kernel using shared memory.

See the concept in real benchmarks

All articles