Skip to content
Skills Library
Programming · Analysis

kernel-profiling-nsight

/kernel-profiling-nsight

Profile GPU work with Nsight Systems and Nsight Compute to read the timeline, name the bottleneck class, and pull the metric that dictates the fix. Use when a GPU program is slower than expected and you need evidence before touching a kernel.

Repository
AI-SKILLS ↗
Licence
MIT
Last repository change
05.09.2026

Assessment and security

100/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content35 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
1
Scripts
0
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 05.09.2026, without references to retired models or APIs.

Licence

MIT, open: use and modify while retaining author attribution.

What it does and when to use it

Profile GPU work with Nsight Systems and Nsight Compute to read the timeline, name the bottleneck class, and pull the metric that dictates the fix. Use when a GPU program is slower than expected and you need evidence before touching a kernel.

Original author description

Profile GPU work with Nsight Systems and Nsight Compute to read the timeline, name the bottleneck class, and pull the metric that dictates the fix. Use when a GPU program is slower than expected and you need evidence before touching a kernel.

How it works

  1. Method
  2. Litmus tests
  3. Boundaries
Original SKILL.md headings · 60 lines, 504 words
SKILL.md preview
---
name: kernel-profiling-nsight
description: Profile GPU work with Nsight Systems and Nsight Compute to read the timeline, name the bottleneck class, and pull the metric that dictates the fix. Use when a GPU program is slower than expected and you need evidence before touching a kernel.
---

# Kernel profiling with Nsight

The two tools split the job cleanly: Nsight Systems shows the whole timeline of
where time goes across CPU, copies, and kernels, while Nsight Compute dissects
one kernel's internals. The common mistake is opening Compute first and missing
that the GPU sat idle 60 percent of the run waiting on the host. Profile
top-down, from timeline to kernel, so the change you make is the one costing
wall-clock time.

## Method

1. **Capture the whole run with Nsight Systems first.** Run `nsys profile -o out
   ./app` and open the report. Scan GPU utilization across the timeline: gaps
   between kernels mean the device is starved by host code, tiny launches, or a
   synchronous `cudaMemcpy`. Close idle time before optimizing any kernel.
2. **Annotate phases with NVTX ranges.** Wrap logical stages (data load,
   forward, backward) in `nvtxRangePush`/`nvtxRangePop` so the timeline reads
   as named bands instead of an undifferentiated wall of kernels. This makes
   the starved phase obvious at a glance.
3. **Name the bottleneck class before optimizing.** Every kernel is capped by
   one of three things: memory bandwidth, compute throughput, or latency from
   too little parallelism. Nsight Compute's Speed Of Light section reports
   Memory and Compute as percentages of peak; high memory and low compute is
   bandwidth-bound, both low is latency-bound and usually low occupancy.
4. **Capture the target kernel with the full set.** Run
   `ncu --set full -k kernel_name -c 1 ./app` to profile one invocation. Filter
   by name and count because profiling every launch is slow, and the full set is
   what lets the guided analysis rules fire.
5. **Read the metrics that name the fix, not raw counters.**
   `dram__throughput` and `sm__throughput` give the SOL split,
   `l1tex__t_sector_hit_rate` shows cache reuse, and low sectors-per-request
   flags uncoalesced loads. Warp State Statistics names the top stall: Long
   Scoreboard is waiting on memory, Barrier is `__syncthreads` contention.
6. **Follow the guided rules, change one thing, reprofile.** Compute prints
   findings like "uncoalesced global access" with an estimated speedup. Treat
   each as a lead, apply a single change, and reprofile, since a fix often
   shifts the bottleneck to a new class rather than removing it.
7. **Diff before and after to prove the win.** Open the two `.ncu-rep`
   files side by side and confirm both the targeted metric moved and
   kernel duration dropped. A counter that improved without cutting
   duration was not the bottleneck.

## Litmus tests

- Can you state memory-, compute-, or latency-bound with SOL percentages behind
  the claim?
- Did you rule out GPU idle time in Nsight Systems before opening a kernel?
- Does the profiler name a concrete stall reason rather than a vague "it's
  slow"?
- After the change, did kernel duration in the report actually fall?

## Boundaries

Nsight tells you where time goes and why a kernel stalls; it does not write the
faster kernel. Turning uncoalesced loads into coalesced ones is
cuda-kernel-basics; turning bandwidth-bound into compute-bound is
gpu-memory-hierarchy. Overhead from `ncu` is high, so profile a reduced input,
and never draw conclusions from a debug build.

Versions

The database has no other skills named kernel-profiling-nsight.

Database snapshot: 13.11.2026 · package version 37237fc96c07 (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 100nsys-optimizer
    Optimize CUDA/GPU simulation code using NVIDIA Nsight Systems (nsys) profiling. Use this skill whenever the user mentions performance problems, slow simulations, profiling, nsys, Nsight Systems, kernel optimization, GPU bottlenecks, or wants to speed up CUDA code. Also trigger…
  • 90kernel-flow
    Use when optimizing a CUDA kernel on an NVIDIA GPU: making a kernel faster, chasing a measured bottleneck with Nsight Compute, or iterating candidates against a correctness matrix. Runs the KDA flow on pi-goal: freeze the shape matrix and tolerances, record a baseline, profile…
  • 100cuda-profiling
    Use when profiling CUDA with Nsight Systems or Nsight Compute, reading roofline and occupancy metrics, or annotating phases with NVTX. Not for correctness: use cuda-debugging.
  • 100perf-nsight-systems
    Nsight Systems (nsys) CLI for system-level timeline profiling. Use when the user wants to run nsys profile, analyze .nsys-rep reports, use nsys stats/analyze/recipe commands, diagnose GPU idle time from timeline traces, or profile distributed training with NCCL overlap analysis.…
  • 100cuda-profiling
    CUDA profiling skill for NVIDIA GPU performance analysis. Use when profiling kernels with Nsight Systems or Nsight Compute, interpreting roofline models, diagnosing memory-bound vs compute-bound kernels, or annotating code with NVTX ranges. Activates on queries about Nsight,…
  • 100nvidia-tensorrt-llm-perf-nsight-systems
    Nsight Systems (nsys) CLI for system-level timeline profiling. Use when the user wants to run nsys profile, analyze .nsys-rep reports, use nsys stats/analyze/recipe commands, diagnose GPU idle time from timeline traces, or profile distributed training with NCCL overlap analysis.…

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.