Definition
Plain language
A tool that measures, in fine detail, what a program is doing inside an NVIDIA graphics chip while it runs.
As stated in the literature
NVIDIA Nsight Compute, a kernel-level GPU profiler reporting counters like occupancy, stall reasons, and throughput; one of the profilers KernelPro's deterministic tools interpret to diagnose bottlenecks before generating optimized code.
Why it matters: It pinpoints exactly where a program is losing time inside the chip, letting engineers fix the real bottleneck instead of guessing.
For example, a developer runs ncu on a slow graphics-chip program and sees that it's spending most of its time waiting on memory rather than computing.
Heard on the show
“" The first, ncu, sees inside the kernel — occupancy, stall reasons, throughput.”Episode 177 — Why Raw Profiler Data Made an AI Worse at Writing GPU Code