Part 9 — Optimization & profiling

You’re a developer, so you know: measure, don’t guess. Performance work is the difference between a tech demo and a shippable game.

9.1 The two budgets

A frame must fit in your target time: 16.6 ms for 60 fps, 33.3 ms for 30. Two parallel budgets:

• Game thread (CPU) — your gameplay logic, ticks, AI, physics setup.
• GPU — rendering: triangles, overdraw, shadows, post-processing, lighting.

Find which one you’re bound by first; optimizing the other won’t help.

9.2 Tools

• stat fps, stat unit (console, backtick ~ key) — unit shows Frame / Game / Draw / GPU ms. This tells you CPU-bound vs GPU-bound at a glance.
• stat GPU — GPU cost broken down by pass.
• Unreal Insights — the heavy-duty timeline profiler (trace a session, analyze CPU/GPU/memory frame by frame). Your “perf flame graph.”
• GPU Visualizer (Ctrl+Shift+,) — per-pass GPU timing.
• Shader Complexity / Quad Overdraw view modes (Viewport → View Mode) — visualize expensive shaders and overdraw as heatmaps.
• stat Game, LLM (memory), obj list — memory and object tracking.

9.3 Common wins (developer instincts apply)

• Kill unnecessary Tick. Per-frame work over many actors is the classic CPU sink. Use timers, events, and overlap callbacks instead. (PrimaryActorTick.bCanEverTick = false by default.)
• Reduce draw calls. Instancing, Static Mesh merging, HLODs. Nanite helps a lot here.
• Mind the lights. Dynamic shadowed lights were historically expensive; MegaLights changes the math, but shadow casting still costs. Cull shadows on small lights.
• Object pooling. Spawning/destroying is costly; pool bullets/effects like any high-churn allocation.
• LODs & culling. Distance culling, occlusion culling, and (for non-Nanite meshes) LODs. Nanite automates geometric LOD.
• Async everything heavy. Load assets async (don’t hitch on LoadObject at runtime); use async traces where possible.
• Profile on the target device, not just your beefy dev machine. A 60 fps editor scene can be 20 fps on a console/handheld.