Definition
Plain language
How much of a GPU's parallel capacity is actually being used — low occupancy means most of the chip is sitting idle.
As stated in the literature
A GPU performance metric for the fraction of available parallel execution slots (warps/blocks) in active use; low occupancy often stems from resource limits like shared-memory pressure, and is a primary bottleneck signal read off profilers like ncu.
Why it matters: Low occupancy is a common reason a GPU runs far below its potential, so it's a key signal for finding where performance is being wasted.
For example, a GPU program that uses too much shared memory per task may leave most of the chip's parallel slots empty, so the hardware sits largely idle.
Heard on the show
“Anyone can read "occupancy is six percent.”Episode 177 — Why Raw Profiler Data Made an AI Worse at Writing GPU Code