Perhaps what we're seeing here is that quantization errors for the non-amplified signal become larger than the noise introduced by the amplification?