Game loop architecture

Game loop architecture

The core decision: simulation runs on a fixed timestep for determinism and stability; rendering runs as fast as it can and interpolates. Mixing them ties gameplay correctness to frame rate, which is how physics breaks on faster machines.

Method

1. Fix the simulation timestep. Update at a constant dt (60Hz or 50Hz commonly): physics integrators stay stable, gameplay is reproducible, and networking/replays become possible (see multiplayer-netcode's determinism needs). Variable-dt simulation makes tunneling, jitter, and speed-depends-on-hardware bugs a permanent background noise.
2. Run the accumulator loop. Each frame: add real elapsed time to an accumulator, run as many fixed updates as it covers, render once with the leftover fraction. The canonical shape: while (acc >= dt) { update(dt); acc -= dt; } render(acc/dt).
3. Interpolate rendering between states. Render entities at prev + (curr - prev) * alpha using the leftover fraction: motion looks smooth at any refresh rate while simulation stays discrete. Store previous transform per entity; snap (no interpolation) on teleports or spawns to avoid ghosting sweeps across the screen.
4. Clamp the accumulator against the spiral of death. A slow frame queues more updates, which makes the next frame slower: cap the accumulated time (e.g. max 5 updates per frame) and drop the excess, accepting slow-motion under overload over a frozen process. Log when the cap hits; it is your overload telemetry (see game-performance budgets).
5. Order the frame deterministically. Input sampling, then fixed updates (gameplay, physics), then late/camera update, then render, then present: one documented order. Input handled inside the fixed step (or queued into it) keeps behavior framerate-independent (see game-input-handling for buffering across the boundary).
6. Keep time sources honest. One monotonic clock for the loop (never wall clock; see clock-skew's lesson locally); pause/timescale implemented by scaling the accumulator feed, not by skipping updates; and separate "game time" from "real time" in APIs so UI animations can run while gameplay is paused.

Boundaries

• Turn-based and event-driven games can use simpler update-on-demand loops; the accumulator machinery earns its keep where continuous simulation exists.
• Engines (Unity/Unreal/Godot) already implement this: there your job is putting logic in the right callback (fixed vs frame update) and respecting their interpolation settings, not rebuilding the loop.
• Frame pacing against the display (vsync, variable refresh) is a presentation concern layered on top; the fixed-step core stays the same underneath.