Kernel Metrics in Nsight Compute

Nsight Compute is the microscope for a single kernel. It collects deep hardware metrics so you can see exactly why it runs slow. 🔬

You profile a kernel with ncu from the command line. It replays the kernel many times to gather detailed counters.

The first thing to read is throughput: how busy the compute units and memory pipes are, shown as a percent of their peak.

A stall is a warp that cannot make progress this cycle. Nsight Compute groups stalls by reason so you know what to fix.

If most stalls say memory, your kernel waits on data. The cure is better access patterns or reuse, not more math.

Achieved occupancy shows how many warps were actually active versus the maximum. Low values often mean too few warps to hide latency.

The roofline plots your kernel against hardware limits. Where the dot lands tells you if compute or memory is the wall.

Nsight Compute writes plain-language recommendations. It points at the biggest bottleneck so you do not have to guess first.

Counters for the L1 and L2 caches show how often reads hit cache. Higher hit rates mean fewer trips to slow global memory.

You can diff two reports side by side. That makes it obvious whether your latest change truly helped the metrics.

Because it replays kernels, profiling overhead is high. Use it on one kernel at a time, not as a normal run.

Your kernel report shows mostly memory-related warp stalls.

Nsight Compute reveals one kernel's metrics: throughput, stalls, occupancy, and roofline. Use it to confirm exactly what to optimize. 👏