Game performance

Game performance

A game's budget is one frame: 16.6ms at 60fps, all-in. Performance work means profiling the frame (not averages), and the enemies are spikes: one 100ms frame reads worse than a steady 40fps.

Method

1. Profile the frame, chase the spikes. Frame-time graphs (not fps averages: fps hides variance), the engine's frame profiler for CPU (which systems, which entities) and GPU (which passes) per frame; capture the bad frames specifically (spike-triggered captures). Diagnose before optimizing: the bottleneck is CPU main thread, render thread, or GPU, and each has different cures (see systems-profiling for the native toolbox underneath).
2. Eliminate steady-state allocation. Per-frame allocations trigger GC pauses (managed engines) or allocator churn (native): pool bullets/particles/effects (pre-allocate, reuse, never destroy mid-play), reuse buffers and collections, avoid closure/boxing garbage in hot callbacks (LINQ and lambdas in Update are the Unity classics; see csharp-linq's allocation note). Target: zero allocations per frame during gameplay, verified in the profiler's GC/alloc view.
3. Cut draw calls with batching and instancing. Thousands of individual draws choke the CPU-to-GPU pipe: static batching for level geometry, GPU instancing for repeated meshes, texture atlases so sprites/materials share draws (see game-asset-pipeline), and UI canvases split so one moving element does not rebuild the whole canvas. Measure draw calls/setpass in the frame debugger before and after.
4. Scale work by distance and visibility. LODs for meshes, culling (frustum plus occlusion where it pays), reduced tick rates for far-away AI/physics (see entity-component-system's system granularity making this easy), and particle/shadow budgets per quality tier. The principle: spend the frame on what the player can see.
5. Spread and defer the spiky work. Level loads and heavy spawns behind async loading screens; procedural generation time-sliced across frames (a budget per frame, resumable work: see procedural-generation); pathfinding and AI queries through request queues with per-frame caps (see backpressure's bounding instinct, in-frame). Hitches at spawn/checkpoint moments are almost always synchronous work that wanted deferral.
6. Gate with performance tests on target hardware. The min-spec device runs a scripted worst-case scene in CI or nightly (see performance-testing, mobile-performance for thermal reality): frame-time p95 and memory ceilings as pass bars, so regressions surface at the commit, not at certification (consoles) or in reviews (mobile).

Boundaries

• Do not micro-optimize before the frame profiler assigns blame; a game is a dozen subsystems and intuition picks the wrong one routinely (see perf-calibration).
• Visual quality trades are design decisions: dynamic resolution, effect density, shadow distance belong in a quality-settings conversation with art direction, not silent engineering cuts.
• Load-time and memory-footprint optimization are their own tracks (streaming, compression: see game-asset-pipeline); frame budget work assumes assets already fit.