AFAIR the NV2A's texture cache is configured in a L1/L2 sort of way. The L2 cache is very small and holds the textures in a compressed form. While the L1 cache is bigger but cannot store compressed textures. The the textures are stored in the smaller L2 cache first and then decompressed from the L2 to the bigger L1 to be worked on.
This isn't all confirmed info though, its bits and pieces from devs on this forum and articles (like the one at Extremetech).