The FFT core accepts complex data samples, but can perform a transform on real-valued data by setting all imaginary input samples to zero.
Due to the finite wordlength effects described previously, noise is introduced during the transform, resulting in the output data not being perfectly symmetric. The DIT and DIF FFT algorithms have different noise effects due to the different calculation order.
For a thorough treatment of this topic, see Limited Dynamic Range of Spectrum Analysis Due To Round off Errors Of The FFT and Influence of Digital Signal Processing on Precision of Power Quality Parameters Measurement.
The asymmetry between the two halves of the result is more noticeable at larger point sizes. In addition, the noise is more prominent in the lower frequency bins. Therefore, AMD recommends that the upper half (N/2+1 to N points) of the output data is used when performing a real-valued FFT.