game-performance

game-performance

Core Philosophy

Game performance optimization is not randomly running Unity's profiler the night before release and guessing which scripts to tweak. Real-time rendering is an unyielding mathematical budget. At 60 frames per second, the entire game engine—CPU game logic, AI pathfinding, physics simulation, audio processing, GPU vertex transformation, rasterization, and post-processing—has strictly 16.66 milliseconds to finish. At 120 FPS, that budget shrinks to 8.33 milliseconds. Hitting frame budgets requires identifying bottlenecks, managing garbage collection allocations, and optimizing draw calls.

---

4-Step Real-Time Performance & Profiling Architecture

Step 1: The Frame Budget Breakdown (16.66ms @ 60 FPS)

1. The Dual-Engine Budget (CPU vs GPU):
   • CPU Budget (Target: $\le 10text{ms}$):
     • Game Logic & Scripts: $\le 4text{ms}$
     • Physics Simulation: $\le 3text{ms}$
     • Render Thread / Draw Call Preparation: $\le 3text{ms}$
   • GPU Budget (Target: $\le 14text{ms}$):
     • Shadow Maps & Geometry Passes: $\le 4text{ms}$
     • Shading & Lighting Passes: $\le 6text{ms}$
     • Post-Processing & UI: $\le 4text{ms}$
2. Diagnose the Bottleneck First:
   • CPU-Bound: GPU utilization is $< 85%$; frame time drops when adding AI entities or physics objects.
   • GPU-Bound: GPU utilization is $99–100%$; frame time drops when increasing screen resolution, shadow quality, or particle density.

Step 2: CPU Optimization & Garbage Collection (GC) Defense

1. Zero-Allocation in the Hot Loop:
   • In Unity / C#, executing new inside Update(), FixedUpdate(), or LateUpdate() allocates heap memory, triggering non-deterministic Garbage Collection spikes that cause visible micro-stutters.
   • Golden Rules:
     • Cache component references in Awake() or Start() (never call GetComponent() in Update).
     • Pre-allocate reusable collections (List<T>, arrays); use list.Clear() instead of new List<T>().
     • Implement Object Pooling for all ephemeral entities (bullets, particles, sound effects, damage numbers).

Step 3: GPU Optimization: Draw Calls, Batching & Overdraw

1. Draw Call Consolidation:
   • Target: Maximum 500–1,000 draw calls per frame on mobile; 2,000–3,000 on PC/consoles.
   • Techniques:
     • Static Batching: Combine non-moving geometry sharing materials into single meshes.
     • GPU Instancing: Render 1,000 identical meshes (trees, grass, projectiles) in a single draw call.
     • Texture Atlasing: Combine multiple prop textures into a single 4K atlas to share a single material.
2. Overdraw & Fill-Rate Defense:
   • Transparent particles (smoke, fire, fog) stacked on top of each other force the GPU to shade the same screen pixel 10 times.
   • Enable Occlusion Culling to prevent rendering objects hidden behind solid walls.

Step 4: Level of Detail (LOD) & Distance Culling

1. The 3-Tier LOD Ramp:
   • LOD 0 (0–15m): 100% mesh triangles (Hero detail).
   • LOD 1 (15–40m): 40% mesh triangles (Simplified bevels, removed micro-props).
   • LOD 2 (> 40m): 10% mesh triangles (Basic silhouette).
   • Billboard / Impostor (> 100m): 2D flat quad facing camera.

---

Deliverable Format: Performance Budget & Profiling Report (PERFORMANCE-BUDGET.md)

# Real-Time Game Performance Budget & Profiling Audit: [Scene Name]
*Target Hardware: Steam Deck / Mid-Spec PC (GTX 1660 / Ryzen 5) | Target: 60 FPS (16.6ms)*

## 1. Frame Time Budget Allocation
| Subsystem | Budget (ms) | Actual Measured (ms) | Status | Primary Culprit |
|---|---|---|---|---|
| CPU Game Logic | 4.0ms | 3.2ms | **Pass** | Clean cached lookups |
| CPU Physics | 3.0ms | 4.8ms | **FAIL** | 200 unbatched rigidbodies colliding |
| CPU Render Thread | 3.0ms | 2.4ms | **Pass** | 820 draw calls |
| GPU Base Pass | 6.0ms | 5.1ms | **Pass** | PBR materials |
| GPU Translucency / FX | 3.0ms | 6.4ms | **FAIL** | Particle overdraw on explosions |
| **Total Frame Time** | **16.6ms** | **21.9ms (45 FPS)**| **FAIL** | Physics + Particle Overdraw |

## 2. Identified Performance Bottlenecks & Fixes
1. **Physics Stutter**: 200 physics debris parts were querying continuous mesh colliders.
   - *Fix*: Switched debris to simple sphere colliders; set kinematic sleep threshold to 0.2s.
2. **Explosion Overdraw**: Smoke particle system spawned 400 overlapping transparent quads.
   - *Fix*: Reduced particle count to 40; used customized cut-out mesh particles to eliminate transparent fill-rate waste.

## 3. Post-Optimization Verification
- **New Total Frame Time**: **13.8ms (72 FPS stable)**. Zero GC allocation in hot loops.

---

Worked Example: Eliminating Garbage Collection Spikes in Unity

• Symptom: Game suffered a 60ms freeze every 8 seconds during heavy combat.
• Profiler Trace: Identified that spawning bullet impacts allocated 45KB of heap per frame, triggering Unity's garbage collector.
• Remediation: Implemented a generic pre-allocated ObjectPool<T> storing 100 reusable bullet impacts.
• Outcome: GC allocations dropped to 0 B per frame in combat; frame rate locked at a solid 60 FPS.

---

Verification Checklist

• Total frame time stays under 16.66ms (60 FPS) or 8.33ms (120 FPS) on target baseline hardware.
• Hot loop methods (Update(), FixedUpdate()) generate strictly 0 B of GC allocation.
• Total draw calls stay within platform limits (<= 1,000 mobile / <= 3,000 PC).
• Object pooling implemented for all high-frequency entities (projectiles, particles).
• Level of Detail (LOD) and Occlusion Culling configured and verified.

---

Anti-Patterns

• Optimizing in the Dark: Guessing what is slow without running a frame profiler (Unity Profiler, Unreal Insights, RenderDoc).
• GetComponent() in Update(): Calling expensive scene hierarchy searches 60 times per second per entity.
• Uncontrolled Alpha Blending: Spawning 1,000 giant transparent smoke particles covering the entire screen.