Procedural generation promises infinite content. Unity DOTS and the Burst compiler deliver the throughput to make it real-time. Most tutorials stop at noise functions. This one shows how to architect a generation pipeline that scales across cores without allocating garbage.
Why DOTS Changes the Math
Classic MonoBehaviour approaches choke on per-chunk overhead. Each GameObject carries transform hierarchy, component lookup, and GC pressure. DOTS replaces that with contiguous arrays of plain structs. The Burst compiler then vectorizes the math — SIMD instructions process four, eight, or sixteen voxels per cycle.
Entity Archetypes for Chunk Data
Define one archetype per chunk resolution. A 32³ voxel chunk becomes a single entity with a blob asset reference. The blob stores the density field — no per-voxel entities. This keeps the entity count low and the chunk count high.
💡
TipUse BlobAssetReference for immutable chunk data. It avoids copy-on-write and enables zero-cost sharing across systems.
Streaming With Job Dependencies
Generation must not stall the main thread. Schedule noise jobs, mesh extraction jobs, and blob baking jobs as a dependency chain. The EntityCommandBufferSystem plays back structural changes at sync points.
Marching Cubes on the GPU Side
Burst handles the density evaluation. For mesh extraction, dispatch a compute shader that reads the blob via a StructuredBuffer. This avoids marshaling vertex data back to CPU. The compute shader writes triangles directly into a GraphicsBuffer for DrawProcedural.
| Stage | Thread | Data Flow |
|---|---|---|
| Noise Eval | Burst (CPU) | NativeArray→Blob |
| Mesh Extract | Compute (GPU) | Blob→GraphicsBuffer |
| Render | GPU | DrawProcedural |
Deterministic Seeds for Multiplayer
Sync the world seed once. Derive per-chunk seeds with a hash: seed = hash(worldSeed, chunkCoord). Every client generates identical terrain without networked voxel data. Only modifications — player-placed blocks — replicate.
"Determinism isn't a feature. It's the absence of hidden state.
— Mike Acton
Memory Budget Guardrails
Allocate a fixed NativeArray pool for active chunks. When the pool fills, evict the furthest chunk using a priority queue keyed by distance. Return its blob to a recycle queue. Zero allocations after warm-up.
⚠️
WarningForgetting to Dispose NativeArrays in the eviction path leaks memory fast. Wrap pools in a IDisposable manager and verify with the Memory Profiler.
Ship It Checklist
- Profile Burst timings with
Unity.Profiling.ProfilerMarkerper job. - Validate determinism: run two clients, compare chunk hashes.
- Stress test: 10k chunks, 60 FPS on mid-tier mobile.
- Add a debug overlay showing job graph latency.
✦
Procedural worlds at scale demand data-oriented design. DOTS and Burst give you the primitives. The architecture — archetypes, blobs, job chains, GPU handoff — is yours to build. Start with a single chunk job. Measure. Extend. The infinite world fits in a finite budget.
