Performance

NV2 Engine is engineered so that GPU bandwidth, draw calls, and CPU meshing stay proportional to what actually changed — not the whole world.

GPU bandwidth: packed vertices

• CPU-side vertices are 72 bytes; the GPU-resident PackedVertex (defined in renderer/vertices.rs) packs them to 36 bytes (Snorm8x4 / Unorm8x4).
• Vertex bandwidth is halved for every chunk upload.

Per-chunk GPU buffers

A block edit re-uploads only the affected chunk instead of the whole loaded area (previously ~81 chunks per change). GPU uploads are additionally debounced rather than re-uploaded every time a single chunk changes.

Frustum culling

Gribb–Hartmann plane extraction culls chunks outside the view, with a safety margin and a protected 3×3 area around the player to prevent over-culling. This reduces visible chunks by roughly half on typical scenes while keeping the player's surroundings always rendered.

Incremental water simulation

• Only changed chunks rebuild water meshes — static water never triggers a rebuild.
• Simulation and meshing are interval-based (water_sim_interval, water_rebuild_interval).
• Water mesh recombination is separated from full water mesh rebuilds.

CPU meshing budget

• Mesh creation is rate-limited — only a small number of chunks are built per frame (mesh_build_budget).
• Seam repair rebuilds only neighboring chunk borders when new chunks arrive.
• Chunk meshes and water meshes are cached separately by chunk coordinate.

Async chunk generation

• Bounded work queue with in-flight deduplication
• Parallel generation via rayon (into_par_iter())
• Results delivered over mpsc — world insertion and mesh coordination stay on the main thread

Background AI training

The neural-network vegetation trainer runs on a background thread:

Performance profiles (settings.rs)

All tunables below are centralized in settings.rs and switched by the Low-End-PC mode (one key in the menus, persisted across runs via settings.json next to the executable).

AppSettings is serialized with serde (pretty JSON), loaded with graceful fallbacks on parse/read errors, and exposed through SharedSettings (an Arc<RwLock&lt;AppSettings&gt;>) for safe concurrent access.

AI implementation impact (measured)