First-pass triage

Establish the bound

1. Run a non-Debug build with the target device profile and resolution.
2. Disable smoothing/VSync/frame caps only when they conceal the bound; record the change.
3. Enter stat unit; add stat unitgraph and stat raw when spikes or filtering matter.
4. Reproduce the workload for a fixed interval.
5. Compare Frame with Game, Draw, RHIT, and GPU.

Interpretation:

Pattern	Working classification	Next capture
Frame follows Game	game-thread bound	CPU/frame/bookmark; inspect Game thread
Frame follows Draw	render-thread bound	CPU/frame/gpu; inspect rendering submission
Frame follows RHIT	RHI/submission/driver bound	CPU/GPU plus platform graphics profiler
GPU stands alone near Frame	GPU bound	profilegpu, GPU trace
all appear below Frame	cap, wait, sync, present, task dependency	timing + context switch; inspect waits
rare raw spikes	hitch	bookmark and narrow hitch capture

stat unit is directional, not final proof. Parallel CPU and GPU work, pipelining, idle time, VSync, and asynchronous compute can make similar numbers misleading.

Fast probes

• Lower primary resolution/screen percentage while holding content constant. A material GPU gain implicates pixel/resolution-scaled work; no gain does not prove a CPU bottleneck.
• Pause or freeze simulation to separate simulation-fed rendering changes from a stable frame.
• Disable one major feature with a supported CVar for diagnosis, then restore it. Treat the delta as attribution evidence, not automatically as the shipping fix.
• Use stat sceneRendering, stat rhi, stat gpu, stat game, stat anim, stat slate, or the subsystem stat only after the first bound is known.