3 ms·

Another factor in favor of 8 bits is that inference speed will generally be 2x faster than a larger model at 4 bits.