NVIDIA Nsight Graphics for UE

Three categories where Nsight is the right answer:

• SM-level performance. SOL%, warp stall reasons (L1 Long Scoreboard, lg_throttle, L2 Limited), L1/L2 hit rates — not exposed by RenderDoc or PIX.
• Ray tracing on RTX. TLAS/BLAS inspection, AABB overlap, RT live state size, Shader Execution Reordering analysis.
• Crash post-mortem. Nsight Aftermath ingests .nv-gpudmp dumps and shows page-fault address → last completed event marker mapping. Required for any shipped title.

Latest version: Nsight Graphics 2026.1.0 (build 26067 Windows x64, 26083 ARM64 Linux). RenderDoc still wins on cross-API and pixel history; PIX wins on Windows D3D12 multi-frame timing. Nsight wins on NVIDIA hardware specifics.

Install Nsight Graphics 2026.1 from developer.nvidia.com. The senior pattern: launch UnrealEditor.exe from Nsight, not attach post-hoc. Hooks fire correctly only at process start.

1. Nsight Graphics → New Project → "Application Executable" → point at UnrealEditor.exe with your project as cmdline.
2. Activity: Frame Profiler for individual frame analysis; GPU Trace Profiler for SM/warp metrics; Generate C++ Capture for offline replay.
3. Click "Launch Frame Profiler" (or equivalent for the activity).
4. Editor boots; play in PIE; capture from the Nsight overlay (F11 default).

Nsight 2025.4+ promoted Graphics Capture from beta — captures persist directly to disk for offline analysis. Workflow:

1. Pre-warm shaders in PIE (UE's PSO compile spikes corrupt timing).
2. Click Capture in the Nsight overlay; capture writes to disk.
3. Open the .nsight-gfx file later for offline analysis.

This is the recommended workflow for CI: a small script can drive a packaged build to a known scenario, capture, and ship the file as an artifact for engineers to inspect.

Use UE's built-in ProfileGPU first to localize the regression to a pass, then Nsight Frame Profiler → "Profile Shaders" on the offending pass for SM-level detail. This two-step is the senior pattern; Nsight first is a waste of capture time when you don't yet know which pass to focus on.

Set r.RDG.Debug 1 + r.RDG.Events 1 for richer event markers so Nsight's nesting matches UE's RDG passes.

GPU Trace Profiler is the deep-telemetry mode. Advanced Mode metrics expose:

• SOL% (Speed of Light). Percentage of theoretical max throughput hit. SM SOL < 50% on a Lumen pass = under-launched dispatch.
• SM Throughput For Active Cycles. Compute density when warps are running.
• L1/L2 hit rates. Memory pressure indicator.
• Warp stall reasons. Why warps aren't issuing — L1 Long Scoreboard (memory wait), lg_throttle (LD/ST throttle), L2 Limited.

The Range Profiler was deprecated; older UE blog posts referencing it are out of date. In 2026.1 the replacement is GPU Trace Profiler (per NVIDIA's migration guide).

For DXR/Lumen HWRT investigations:

• Inspect TLAS/BLAS — instance counts, BLAS overlap, build flags.
• UE5 emits these via RDG; require r.RayTracing=1 and Lumen HWRT before launching Nsight.
• RGP-equivalent BVH analysis is in the Ray Tracing Inspector tab.

NVIDIA's Indiana Jones path-tracing case study reported an order-of-magnitude traversal-cost improvement after splitting misaligned BLAS — the exact pattern Nsight surfaces (per NVIDIA's case study).

SER is RTX 40-series's hardware-accelerated reordering of divergent rays into coherent groups. Open the Ray Tracing Live State tab in GPU Trace; identify GLSL/HLSL declarations spilled across traceRay/callable boundaries.

From the Indiana Jones case study: SER alone 4.08 ms → 3.63 ms (~11% gain); with RT live state reduced from 222 to 84 bytes, SER yielded 4.07 ms → 3.08 ms (~24% gain); active threads/warp went 38% to 68–70%.

UE 5.6 enables SER by default for Lumen HWRT (r.Lumen.HardwareRayTracing.ShaderExecutionReordering=1) on supported hardware. The NVIDIA-published number for Lumen reflections in CitySample is 20–30% faster with SER on RTX 40 (per NVIDIA's SER announcement).

For shipping titles, Aftermath is non-negotiable. Crash dumps from production let you find the offending warp without a repro:

1. Build with the Aftermath SDK integrated (UE 5.3+ docs cover this; SDK 2023.2+ recommended).
2. Run with -gpucrashdebugging command-line. (Do not combine with -d3ddebug — UE docs explicitly warn this combination produces unusable output.)
3. On crash, .nv-gpudmp file written to Saved/Crashes/.
4. Ingest in Nsight Graphics 2023.3+: page-fault address → last-completed event marker; SM register dump → offending warp.

Configure r.D3D12.GPUTimeout to be more permissive for long Aftermath captures — default Windows TDR cuts the dump short.

For non-NVIDIA hardware (AMD, Intel), reach for RGP / Radeon tools or RenderDoc. Nsight Graphics is NVIDIA-only.

PerfGuard regression reports flag pass-level GPU timing deltas; the natural drill-down for an NVIDIA-specific RT regression is opening that scenario in Nsight's Ray Tracing Live State tab. PerfGuard CI can drop .nv-gpudmp and capture artifacts next to its baseline JSON so the engineer arrives with the data already loaded.

• RenderDoc for Unreal — cross-API alternative.
• PIX on Windows — D3D12-specific timing capture.
• AMD Vendor Tools — the AMD analog (RGP / RRA / RMV).
• Lumen Performance — the system most often investigated in Nsight.