Is Global Memory Access Cached Or Not?

Hello all,

One main conclusion I took home from this year CUDA workshop in Dresden was that memory access to global memory is NOT cached.

Now I read a paper from Govindaraju, et al. (see here), stating that the G80 has a Cache size of 392KB.

So memory access to global memory is cached after all?

Any insights?

Thanks and best regards,

maybe he’s referring to textures. They’re read from global memory, but pass through the texture cache.

Texture cache is much smaller than that AFAIK.

392 is summed across all MPs and across all cache types. There’s a lot of different caches in a GPU, many we can’t access (like the one for framebuffering and antialiasing), some we can’t access fully, and all we can’t access in entirety from just one MP.