Re:but we already have directx (Score 1)

Regarding your statement: "In fact most modern hardware (Radeon, Geforce) will prefer it if you pack all your vertex components together, as DirectX has always worked. This all to make more efficient use of memory bandwidth. "

For GeForce this is NOT true. The cache for DMA transfers of non-interleaved arrays makes it just as good as interleaved arrays (or packed components). The primary exception here would be if padding were necessary for an individual array (eg GL_SHORT normals).

In general, it's much more important to write sequentially to AGP memory since the memory is uncached. This usually means that static vertex attributes should be *not* be interleaved with dynamic ones.

Thanks -
Cass