An application-learnable neuromorphic frequency synthesizer with random vector neural network

This paper proposes a novel architecture of application-learnable neuromorphic frequency synthesizer (ALNFS). We train a feedforward random vector neural network (RVNN) to obtain weights. Then by shifting trained weights, ALNFS can operate as both Integer-N and Fractional-N synthesizing, with 0 cycle frequency switching. With 50 hidden neurons, the prototype ALNFS achieves -68.71 dBc/Hz Phase Noise at 100 KHz offset of 4 MHz output frequency while power consumption is only 3.59 μW for RVNN.