horizon/docs/media/entity-culling-benchmark.md

Entity culling benchmark

Run date: 2026-11-01T14:09:25Z.

Setup

• Hardware: Intel Core i7-3770S (8 threads); Intel HD Graphics 4000 (IVB GT2) with Intel Mesa 26.2.4-arch1.1; Vulkan; 1366x768 @ 60 Hz; Linux / Hyprland Wayland.
• Build: cargo build --profile fast -p horizon-client (successful).
• Scene: 300 static armor stands behind a 33-wide x 31-high stone wall at 3.5 blocks from camera, with a 1-wide x 4-high center opening.
• Same seed-1234 scene per pass; 1366x768 fullscreen, render distance 2, immediate presentation, VSync off, 260 FPS cap.
• Camera: Fixed at the initial spawn X/Z, feet y=100, yaw 0 and pitch 0. Time was set to 6000 at scene setup and then advanced normally.
• Six 10 second samples per mode across both orders: Off → Sync → Async, then Async → Sync → Off. At least 5 seconds to settle after changes.

Median results

Mode	Average FPS	Change vs Off	1% low FPS	CPU average	P99 frame time
Off	83.59	—	60.48	8.66 ms	14.99 ms
Sync	96.90	+15.9%	62.83	6.82 ms	12.99 ms
Async	99.67	+19.2%	64.55	5.99 ms	13.02 ms

Individual samples

Pass	Mode	Average FPS	1% low FPS	CPU average	P99 frame time
Forward	Off	83.37	46.68	8.88 ms	19.27 ms
Forward	Sync	96.55	62.56	7.26 ms	13.52 ms
Forward	Async	103.77	81.52	5.83 ms	11.93 ms
Reverse	Async	98.55	45.93	6.17 ms	18.66 ms
Reverse	Sync	97.26	56.59	6.45 ms	13.04 ms
Reverse	Off	83.98	57.35	8.60 ms	14.60 ms

GPU timing was unavailable because this driver returned zero for all timestamp queries. This is one synthetic scene on one Intel HD 4000 system; results are workload-specific.