Definition
Plain language
The actual machine instructions a GPU was told to run, read back by disassembling the compiled program — the ground truth of what the hardware really did.
As stated in the literature
NVIDIA's low-level GPU assembly; disassembling to SASS lets analysts verify what instructions actually compiled (e.g., catching silent scalar fallbacks where tensor-core instructions were expected) when higher-level metrics can't disambiguate the cause.
Why it matters: It reveals what the hardware actually did, so engineers can catch silent performance fallbacks that higher-level metrics would never expose.
For example, an analyst disassembles a compiled GPU program to confirm it really used the fast matrix-multiply instructions instead of silently dropping back to slow ones.
Heard on the show
“It's called SASS, and it disassembles the actual compiled machine instructions — what the hardware was literally told to do.”Episode 177 — Why Raw Profiler Data Made an AI Worse at Writing GPU Code