glossary/tooling/nsight
Nsight
also Nsight Systems, Nsight Compute, nsys, ncu
NVIDIA's profilers. Nsight Systems records a timeline of CPU calls, kernel launches and memory transfers, which shows whether a step is bound by the GPU or by the host feeding it. Nsight Compute profiles one kernel in depth against the roofline, showing achieved bandwidth and occupancy. Together they replace guessing about why a step is slow with measurement, and every claim on this site about what a phase is bound by is the kind they can check.
one timeline
Of launches and gaps, the fastest way to see launch overhead or a starved GPU.
Related
- glossary/
- Kernel launch overhead
The fixed cost of asking the GPU to run a kernel: the driver call, the scheduling, the queue.
- Roofline model
A chart with arithmetic intensity on the x axis and attainable performance on the y, bounded by a sloped line from memory bandwidth and a flat line from peak compute.
- CUDA graphs
Recording a sequence of kernel launches once and replaying the whole sequence with a single call.
- Memory-bound
Limited by how fast bytes can be read, not by how fast they can be operated on.
- sources/
- topics/