
上QQ阅读APP看书,第一时间看更新
Cache memory
On the latest GPUs, there is an L1 cache per multiprocessor and an L2 cache, which is shared between all multiprocessors. Both global and local memories use these caches. As L1 is near to thread execution, it is very fast. As shown in the diagram for memory architecture earlier, the L1 cache and shared memory use the same 64 KB. Both can be configured for how many bytes they will use out of the 64 KB. All global memory access goes through an L2 cache. Texture memory and constant memory have their separate caches.