Wookash Podcast Can Fiddling With Meshes Improve Game’s Performance?
Mar 21, 2026
Arseny Kapoulkine, a software and graphics engineer behind MeshOptimizer and gltf tools. He breaks down mesh vertex quirks, cache-friendly index ordering, meshlet and mesh shader workflows, mesh compression versus runtime cost, quantization trade-offs, and Nanite-style cluster LOD approaches. Short explanations of tools, metrics, and practical pipeline choices.
AI Snips
Chapters
Transcript
Episode notes
Index Order Matters Because Of Vertex Reuse Heuristics
- Vertex cache behavior is the main historical reason index order matters: GPUs reuse previously shaded vertex outputs via small caches or per-batch deduplication.
- MeshOptimizer uses a heuristic Forsyth-style reordering tuned by offline optimization of a small table of weights across many meshes and simulated hardware models.
Follow The Minimum Mesh Optimization Pipeline
- Minimum pipeline: reindex duplicates, optimize index order for vertex cache, optimize vertex fetch, then quantize vertex attributes before shipping runtime data.
- Quantize normals/tangents aggressively but treat positions and UVs conservatively and test per-asset or by category.
Quantize Direction Vectors Heavily But Test Positions
- Quantization trades size/ bandwidth for visual error and must be tuned per attribute; normals tolerate more compression than positions and UVs.
- Many engines safely quantize direction vectors (normals/tangents) to ~20–24 bits using octahedral encodings with minimal artist complaints.