Any modern graphcis card will have some form of cache but it's not even remotely analogous to a microprocessors cache hierarchy/design or function. The usage of L1/L2 cache is not applicable for graphics cards, it's neither useful nor required as it is for a typical generic purpose processor like the X86 P4/Athlon.
For graphics cards one may as well think of main memory as it's cache... but if your referencing SRAM then the largest quantities of cache you'll find is texturization internal cache.... which generally ranges from 1-8K in size in mainstream graphics cards. It's used for texels from the texture maps as they are read from the frame buffer.
The R8500 specifically has 2K of texture cache, while the R9000 has that bumped to 4K.
The triangle setup engine also contains a very small amount of internal cache primarily to store indexed strips, with indices referring to vertices.
There is also the pixel cache used to cache pixels that need to be written back to the frame buffer.
Havent a clue as to the size of the pixel cache.... I think the old Rage 128 had 8K pixel and texture cache though.
I don't think nVidia has ever publically revealed the cache any of their graphics chips use.
I'm not sure if anything I've said here even makes sense to you.... but then I'm hardly the most qualified person to answer this question.
Then again, I'm not sure if there are any around that can give you a much better explanation so hopefully this answers your question somewhat,