This is true but the instruction already existed and it doesn't support uint16_t accumulation. For the reason you mention, activations are uint8_t and weights are int8_t so it worked out well for neural networks.