The overhead of converting to and from floating-point arithmetic is high, so the graphics instructions are optimized for short-integer arithmetic. Image components are 8 or 16 bits. Intermediate results are 16 or 32 bits.