VIVIA — Performance
Benchmarks
Benchmark targets and methodology are specified in
BENCHMARKING.md
(frame-time variance, voxel meshing latency, rendering throughput). The Criterion
harness implementing them is specified there and lands in the repo before launch —
until then, treat the numbers below as targets, not measurements.
Targets
| Metric | Target |
|---|---|
| Average frame time | < 16.67 ms (60 FPS) |
| 99th percentile frame time | < 33 ms (30 FPS floor) |
The engine is built to hit these budgets: no per-frame allocations on the hot path, rayon-parallel chunk generation off the render thread.
MeMLP Inference
| Operation | Cost |
|---|---|
| Forward pass | sub-0.1 ms (~1,100 parameters, pure ndarray math) |
| Training epoch | single-digit ms per 100 samples (background thread) |
| Checkpoint | ~14 KB JSON |
The AI runs on a background thread — training never blocks the frame loop.
Memory
| Component | Size |
|---|---|
| MeMLP checkpoint | ~14 KB (JSON) |
| Per-chunk storage | 128 KB (16×512×16, 1 byte/block) |
| World mesh cache | bounded, invalidated on block changes |
Scalability
| Metric | Value |
|---|---|
| Chunk height | 512 blocks (double the standard voxel height) |
| Render distance | configurable load_radius / render_radius / cleanup_radius |
| Low-end mode | low_end_pc profile reduces radii for modest hardware |
| World size | unlimited (procedural, seed-driven) |
The honest footnote
Every number on this page is either a target (frame times, to be measured by the harness) or an approximation on dev hardware (Arch Linux, mid-range GPU). The harness spec exists precisely so every claim can be reproduced — and the store page will only publish numbers measured by it.