Reference [4] recommends that the IFFT size \(N_{FFT}\) should be 10 times the length of the radar pulses and a power of two. The radar pulses in the GOTCHA data set each have ~424 samples and so a baseline size of \(N_{FFT}=4096\) seems reasonable for the baseline MATLAB system model. This brings challenges for AI Engine implementation because 4096 x 8 Bytes = 32 KB of storage required for a single input or output buffer. Because the memory footprint for the SAR engine is quite large (as outlined in the System Partitioning section), we adopt \(N_{FFT}=2048\) for the AI Engine implementation to somewhat alleviate this concern. The radar pulses from the GOTCHA data set are zero-padded from ~424 samples to match \(N_{FFT}\).