3 ms·

I assume he means quantization (e.g. scaling the weights from 16-bit to 4-bit) and it speeds up the output by reducing the amount of work done.