The depth buffer is spent next to the camera.

At near = 0.1 and far = 10000 — the two numbers everybody types — 99.9% of a 24-bit depth buffer is consumed by the first one percent of the view range, and half of it is gone by 0.2 units. So when the distance z-fights, the instinct is to pull the far plane in. At 1000 units out, lowering far ten-fold is worth 1.00009×. Raising near ten-fold is worth 10× — everywhere.

The engine is free and needs no account. Paste the scene — one camera, the canvases it renders into, the objects in front of it and the lights that cast from it. fov is the VERTICAL angle, so the aspect decides the horizontal shot: one 50° camera sees 73.5° across on a 1440×900 canvas and 24.3° on a 390×844 one, while the vertical never moves — and there is no horizontal field of view to set, only the one the canvas gives you. A shadow map is a second frustum with its own near, far and texel density of mapSize ÷ span, so halving the span is worth exactly as much as doubling the map and costs nothing instead of four times the memory. And culling tests a bounding SPHERE, not the object: a 100×100 ground plane is submitted to the GPU from 20.7 units outside the frustum, every frame, and nobody can see it. Every lane ships a worked example that costs nothing. Writing the report needs credits.