perf-tuning

oleh nvidia

Mengatasi masalah kinerja umum Kit/Isaac Sim/Isaac Lab menggunakan pengaturan dan perubahan konfigurasi tertentu. Mencakup hambatan PresentFrame,…

npx skills add https://github.com/nvidia/omniperf --skill perf-tuning

Performance Tuning for Kit / Isaac Sim / Isaac Lab

Specific fixes for performance issues identified through profiling. Prerequisite: you should already know where the bottleneck is (from diagnose-perf, profiling, or nsys-analyze skills).

PresentFrame is Abnormally Slow

Two causes:

GPU Backpressure

GPU work exceeds frame budget → CPU waits for swapchain buffer.

Verify: Enable GPU zones in Tracy and check if GPU frametime > CPU frametime.

--/profiler/gpu/tracyInject/enabled=true --/profiler/gpu=true --/rtx/addTileGpuAnnotations=true

Fix: Reduce GPU workload — see RTX Tuning below.

VNC / Remote Desktop

Virtual framebuffer causes driver present timing issues.

Fix: Use a physical monitor or run in headless mode.

resolveSamplerFeedback is Abnormally Slow

Texture Streaming bug causes per-frame thread waits proportional to RenderProduct count.

Fix: Disable Texture Streaming:

--/rtx-transient/resourcemanager/enableTextureStreaming=false

Impact: ~6.72 ms saved per frame (measured). VRAM usage may increase — verify headroom.

Headless Mode (3–4 ms GPU savings)

For simulation-only workloads (RL training, SDG, automated benchmarks):

# Isaac Sim via SimulationApp
simulation_app = SimulationApp({"headless": True, "disable_viewport_updates": True})
# Kit args
--no-window --/app/window/hideUi=True

Also disable viewport updates in Python:

from omni.kit.viewport.utility import get_active_viewport
get_active_viewport().updates_enabled = False

Multi-Camera Render Count Verification

Always check Tracy GPU zones to verify only intended cameras are rendering:

  • Count extrt/rtx/rtaTexturesMC_* (camera textures) and t_viewport_ViewportTexture_* (viewport textures)
  • Verify resolution in zone names matches intent (e.g., rtaTexturesMC_3_RP_1920x1080)
  • In headless mode, check that unnecessary viewport textures aren't rendering in background

Common Multi-Camera Fixes

  • Remove per-camera viewports; keep render products only.
  • Replace separate camera render products with TiledCameraSensor when the workflow supports it.
  • Destroy or disable the default viewport after sensor setup in headless benchmark/simulation runs.
from isaacsim.sensors.experimental.camera import TiledCameraSensor

sensor = TiledCameraSensor(
    camera_paths,          # List[str]
    resolution=(H, W),     # Height, Width
    annotators=["rgb"],
)
data, info = sensor.get_data("rgb")

PhysX Tuning

Expose Full PhysX Detail

Default profilerMask=1 hides internal PhysX zones. To diagnose PhysX bottlenecks, remove the mask arg (defaults to ALL).

Key Settings

SettingEffect
--/physics/suppressReadback=trueSuppress GPU→CPU readback
--/physics/updateToUsd=falseSkip physics→USD writeback (only if not reading state from USD)
--/physics/disableContactProcessing=trueSkip contact event callbacks (if no subscriber needs them)

Async Physics (USD Schema)

from pxr import UsdPhysics, PhysxSchema

stage = omni.usd.get_context().get_stage()
for prim in stage.Traverse():
    if prim.IsA(UsdPhysics.Scene):
        api = PhysxSchema.PhysxSceneAPI.Apply(prim)
        api.CreateUpdateTypeAttr().Set(PhysxSchema.Tokens.asynchronous)

Applicable when previous-frame physics results are acceptable (RL training, SDG).

Solver Type

The default solver is PGS. The profiling guide calls out switching PhysxSceneAPI.solverType to TGS as a scenario-dependent tuning option; verify with a WARM benchmark before keeping it.

api = PhysxSchema.PhysxSceneAPI.Apply(physics_scene_prim)
api.CreateSolverTypeAttr().Set("TGS")

Extension Change Detection (fsWatcher)

~0.1 ms/frame overhead for hot-reload file monitoring. Disable for benchmarks:

--/app/extensions/fsWatcherEnabled=false

Viewport Gizmo Overhead

Gizmos (manipulators, grid, selection outlines) cause significant CPU overhead in scenes with many objects.

--/persistent/app/viewport/displayOptions=0
--/persistent/app/viewport/gizmo/enabled=false

DLSS-G (Frame Generation) — Know What You're Measuring

DLSS-G inserts AI sub-frames between real frames:

  • Viewport HUD FPS includes sub-frames → inflated by 2–4x
  • Tracy/benchmark frametime FPS = actual rendering FPS
GPU GenSub-frame Multiplier
Ada (RTX 40)2x
Blackwell (RTX 50)3–4x

Rule: DLSS-G is for viewport display only. Disable for simulation/benchmarks:

--/rtx-transient/dlssg/enabled=false  # default is off, but verify

HydraEngine waitIdle

When /app/hydraEngine/waitIdle=true (default), main thread blocks every frame until GPU finishes.

--/app/hydraEngine/waitIdle=false  # allow CPU-GPU pipelining

Safe for rendering-only workloads. Keep true if reading GPU results same-frame (physics readback, synchronous sensors).

Multi-GPU — Not Always Faster

Multi-GPU adds CPU overhead (job distribution, per-GPU setup, data gathering). Only effective when GPU is the clear bottleneck.

ScenarioRecommendation
High-res (4K) × many cameras (4–8)Multi-GPU effective
Low-res (1080p) × few cameras (1–2)Single GPU faster
CPU-bottleneckedAdding GPUs = pointless

Always verify GPU-bound with Tracy/nsys before adding GPUs.

CPU Governor

# Check
cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor | sort | uniq -c
# Fix
echo performance | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

Real impact: powersaveperformance saved ~4 ms/frame in measured cases.

RTX Tuning When GPU-Bound

Recommended Approach

Start from an IsaacLab preset, then selectively re-enable features you need.

IsaacLab Presets (ChessRTX path-traced, RTX PRO 6000 Blackwell)

PresetFPSΔ vs default
(default)51.85
isaaclab_quality70.25+35.5%
isaaclab_balanced87.81+69.4%
isaaclab_performance111.90+115.8%

DLSS execMode — Single Biggest Lever

execModeModeFPSΔ vs default
0Performance106.24+104.9%
1Balanced87.08+68.0%
2Quality71.42+37.7%
--/rtx/post/dlss/execMode=0  # always specify 0/1/2 — mode 3 (Auto) broken in headless

~91% of isaaclab_performance's gain comes from this single setting.

Full Preset Parameter Table

For detailed per-parameter comparison across all three presets, see references/rtx-presets.md.

Tuning Checklist

  1. Confirm GPU-bound with Tracy/nsys
  2. Verify DLSS-G off for simulation (--/rtx-transient/dlssg/enabled=false)
  3. Set --/rtx/post/dlss/execMode=0 (biggest single lever)
  4. Apply isaaclab_performance preset for maximum speed
  5. Selectively re-enable visual features as needed
  6. Scene-dependent — numbers will vary

Lebih banyak skill dari nvidia

compileiq-debug
nvidia
Gunakan ketika ada yang salah: Search() menggantung, semua evaluasi mengembalikan INVALID_SCORE, skor tidak kunjung membaik, setiap konfigurasi mengembalikan angka yang sama, error ptxas…
create-github-pr
nvidia
Buat pull request GitHub menggunakan gh CLI. Gunakan saat pengguna ingin membuat PR baru, mengirimkan kode untuk ditinjau, atau membuka pull request. Kata kunci pemicu -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Memindai isu terbuka lainnya untuk menemukan isu yang mungkin juga diperbaiki atau secara tidak sengaja dirusak oleh suatu PR tertentu. Menghasilkan peluang perbaikan yang berdekatan dan risiko kontradiksi dengan file:baris…
fhir-basics
nvidia
Mengajarkan agen cara kerja API FHIR R4, sumber daya apa saja yang tersedia, cara melakukan kueri dengan parameter pencarian, dan cara mengurai semua format respons dengan benar…
compileiq-validate-result
nvidia
Gunakan SETELAH Pencarian selesai dan SEBELUM mengklaim percepatan atau mengirim ACF. Muat CSV dump_results, ekstrak kandidat top-K (tujuan tunggal)…
changelog-audit
nvidia
Audit Warp CHANGELOG.md sebelum rilis: pulihkan entri yang hilang, urutkan berdasarkan dampak pengguna, perbaiki bahasa entri, bungkus baris, dan (mode cabang rilis) naikkan bandingkan…
maintain-dynamic-plugins
nvidia
Mempertahankan pemuat plugin dinamis NeMo Relay, manifes, SDK asli Rust, protokol pekerja gRPC, SDK pekerja Python, dokumen, pengujian, dan cakupan alur kerja rilis
dgx-diagnose
nvidia
Diagnosis masalah umum DGX Station GB300 — crash CUDA, penargetan GPU yang salah, bug kontainer vLLM/SGLang, masalah status MIG, kesalahan NVLink/Fabric Manager,…