What the Analog FFT (AFFT) can do

The SiliconIntervention AFFT performs three things:

  1. It acquires the signal (from Lidar or Audio or many other applications) and
  2. Finds the frequencies present in that signal before presenting the data to the AI or controlling DSP.
  3. Vastly reduces the data rate of signals going to the AI/DSP so saving weight and power while reducing physical size.

It is valuable because it does these things at very much lower power and hundreds of times faster than the combination of conventional ADC and AI/DSP can do.

The Principle of Operation

The Analog FFT computer (AFFT) is a fixed-weight neural network of analog computations that performs a Fourier analysis – it breaks the signal down into its frequency components. It is an example of dynamic aggregate behavior.

It is a neuromorphic analog computer exploiting functionality that emerges from complexity of interconnection: this is what we do at SiliconIntervention. This is a quarter of a million analog devices configured and laid out by code to perform a remarkable and commercially valuable signal processing function.

Way beyond an op-amp or DAC/ADC, this is the new analog in action. It’s the analog design paradigm that works on 2nm processes.

On a digital-only advanced CMOS process (it uses inverter-like amplifiers and needs only MOM or MIM capacitors) the AFFT can calculate a Fourier transform to 128 bins in <1uS to >70dB SFDR consuming only 900uW.

The same circuit architecture biased for higher speed can calculate a 128 bin FFT in 2.8nS on a 0.15u process consuming 150mW. On 22nm or lower GigaHertz rate FFT calculations can be achieved.

The same architecture biased for lower power can calculate Audio Rate (48kFFT/sec) FFTs consuming significantly less than 1 micro-Watt.

This is helpful because typical data acquisition systems such as voice recognition have a conventional Analog to Digital Convertor (ADC) followed by a digital signal processor (DSP) to find the Fourier transform of that signal.

The electronic circuit is created from code: the well-known Cooley-Tukey Radix 2 Decimation in time algorithm is run with a programmable number of bins (N) and outputs a netlist corresponding to a log2(N) deep layer neural network (NN) each layer having 2N neurons.

The resulting NN may be activated as a round-robin sampler, and since the layers are analog computations, the first layer can directly sample the input signal, replacing the ADC: no ADC is needed to load the NN.

In a breakthrough innovation the SiliconIntervention AFFT architecture can achieve a bin-spreading in the output bins of the AFFT equivalent to an advanced Window Function despite having an equally weighted round-robin input sampler. This means the AFFT can either:

  • Sample at mutli-gigahertz speeds for Radio and Lidar applications


  • Achieve very low (nano-Watt levels) in Audio applications

We are currently developing a complete analog to MFC convertor that performs data sampling, FFT calculation, bin spreading suppression, logarithmic absolute value calculation, summation into a perceptually uniform audio space (MELS), a second FFT of that resulting MELS space (hence a Cepstrum) , encoding into multiple digital outputs at millisecond rates for the AI/DSP and capable of applying a custom interrupt generator from an arbitrary pattern of ceptrum bins; all while consuming tiny micro-watt levels of power.

This is by far the lowest power and lowest data rate means to construct an always-on Voice Activity Detector, Pulse Oximeter, or Sound Environment Analysis system.

Contact us for even more details.

Learn more about static aggregate behavior: an example to find the minimum and maximum of a signal.