NVIDIA-Omniverse Performance Triage
Use this skill as the entry point for any "my Kit/OV/USD app is slow" question. It does three things:
- Apply the three first-step checks before any deep diagnosis.
- Classify the symptom into a known bucket.
- Route to the specialist skill that owns the bucket.
Step 1 — first-step checks (always do these first)
These three checks resolve a surprising fraction of "performance issues" outright. Apply them before any deeper diagnosis.
- Reproduce against the latest Kit. Defaults assume Kit 109 / 110. Kit 107 is end-of-life or near it. Many issues are already fixed upstream.
- Reproduce against a stock Kit App Editor template with custom extensions disabled. If the issue does not reproduce, the problem is in the user's customizations, not Kit. Repo: https://github.com/NVIDIA-Omniverse/kit-app-template.
- Check the GPU driver against Technical Requirements. Wrong/unsupported drivers are one of the most common single root causes for crashes and surprising perf issues.
Step 2 — classify the symptom
Match the user's description to one of these buckets:
| Symptom | Bucket | Next skill |
|---|---|---|
| Stage takes a long time to open | Loading | profiling-guide (activity timeline), then usd-scene-structure and ujitso-manager |
| Low FPS during navigation | Frame time | profiling-guide, then either rendering-tuner (GPU-bound) or usdrt-advisor (CPU-bound) |
| Excessive memory usage / OOM | Memory | memory-optimizer, then fsd-configurator if instanced |
| GPU crash / "device-lost" | Stability | troubleshooting-router |
| Performance regression after Kit upgrade | Regression | troubleshooting-router (Scenario 6) |
| "How do I structure my USD" / authoring questions | Authoring | usd-scene-structure |
| "How do I import CAD" / content pipeline | Content prep | content-prep-advisor |
| Cloud / container / cold-start | Deployment | cloud-perf-advisor |
| Work is headless / standalone OpenUSD — fix the file, no running Kit app needed | Authoring / optimization | Hand off to the USD Performance Tuning skill for an agent-driven optimization pass (see AGENTS.md → "When to hand off to the USD Performance Tuning skill") |
| Workload is Isaac Sim or Isaac Lab specifically | Robotics / RL | Hand off to Omniperf — start with diagnose-perf, then chain into Omniperf's profiling and nsys-analyze |
| Need to install Nsight Systems / Tracy / capture tools | Tooling | Omniperf install-profilers |
| Need a measured FPS / GPU-util / startup number for comparison | Reference data | Omniperf benchmark dashboards |
Step 3 — what to do if the question is mixed
A typical mixed question (e.g., "my facility scene is at 5 FPS, has 300k instances, and 2000 materials") chains skills:
perf-triage
→ profiling-guide (confirm GPU- or CPU-bound)
→ usd-scene-structure (verify instanceable=true is on referenced/payloaded prims)
→ fsd-configurator (mergeMaterials, enableRendererInstancing — caveats apply)
→ rendering-tuner (DLSS, bounce counts on glass)
→ scene-optimizer-guide (only if the above don't resolve)
Always do triage first. Don't jump to Scene Optimizer or settings tuning before profiling.
Defaults to recommend
- ✅ Run the latest Kit.
- ✅ Use binary
.usdcfiles; reserve.usdafor small interface/debug files. - ✅ Keep
omni.kit.profiler.windowandomni.kit.profiler.tracyextensions loaded so F8 and Tracy actually work. - ❌ Don't tune render settings before knowing it's a rendering problem.
- ❌ Don't add GPUs to fix a CPU bottleneck.
Things to never recommend without qualification
- Specific FPS targets, prim-count thresholds, % improvements, or multi-GPU scaling factors. None of those are citable.
- The deprecated setting
/app/usdrt/population/utils/instanceCompactTransforms(deprecated since Kit 108).
Source guide section
docs/get-started/quick-start.md, docs/get-started/profiling.md, docs/workflows/troubleshooting.md.