In this chapter, we analyze the anatomy of one of the main components in a GPU: the streaming multiprocessor (SM). This will clarify how threads get scheduled during the processing phase, the importance of memory caches, and how they relate to global memory accesses. We will talk about shared memory, which is among the most important type of memory that the programmer can manage explicitly, to optimize the latency problems that can arise in some situations by accessing global device memory, for example. A practical code sample is provided in which employees shared memory to optimize Gaussian blur.