Generated by objanalyze.

GPUs have what is called a post-transform vertex cache. One day, I'll explain it plainly, but for now I'll just link to ryg's thoughts on the matter.

ACMR (average cache miss ratio) is the number of cache misses per triangle. The best this can be is 0.5, but it depends on the mesh; for example, boundaries and seams in a mesh increase ACMR because of reduced vertex sharing. ATVR (average transform to vertex ratio) is the number of cache misses per vertex. The best this can be is 1.0, and is independent of mesh. Ignacio Castaño explains this in this blog post, but the gist is that ACMR is more common, and ATVR is more useful because it is less sensitive to a particular model.

hand_00.obj

9740 vertices, 17135 triangles

Before:

After:

ben_00.obj

44915 vertices, 78029 triangles

Before:

After: