View article

We present inexact Fast Fourier Transform (FFT) accelerators that can realize energy-accuracy tradeoffs taking advantage of various inexact design techniques in conjunction with a machine-learning inspired waveform shaping technique. A 65nm ASIC test chip with several inexact FFTs shows a reduction in datapath energy consumption upto 75% and the total energy consumption upto 45% (a factor of 4X and 1.8X respectively) when compared to a conventional exact FFT at marginal Signal-to-Noise Ratio (SNR) losses between 0.0002 dB to 1 dB.