High-level architecture of the NVIDIA GTX 480 graphics card (simplified and excluding the graphics-specific details).

<p>The GPU contains several multiprocessors (MPs) for executing groups of threads, called “thread blocks", assigned to MPs by the global control logic (scheduler). Each MP contains several (2×16 for GTX 480) CUDA cores capable of performing floating-point multiply-and-add and integer, logical and bitwise operations. MPs also contain a fast programmable shared memory, hardware-managed data caches and “special function" units that perform double-precision or more complex floating-point operation (such as reciprocal, square root, sine and cosign). A local scheduler assigns resources (including CUDA cores) to the thread blocks assigned to the MP.