An All-Digital Synthesizer Enabled by a Convolutional Neural Network

An “All-Digital” synthesizer is proposed comprised of an inverter-based vector modulator and a Convolutional Neural Network (CNN) controller - all integrable within a digital CMOS process. The modulator is broadband tunable (GHz's of bandwidth), broad phase range (≈ 100°), and high resolution (< 0.1°) enabling precision tracking. Each channel of the synthesizer has an expansive control space, in the order of 2²⁸ states, and exhibits a highly non-linear response to the control code. A vector mapping encoding is proposed to characterize the control state-space by a CNN producing a 1.44° RMS prediction error with only 0.01 % of the input state-space as a training set. A four-channel modulator was fabricated in a 12 nm FinFet CMOS process achieving an operating range of 4 GHz to 8 GHz, an Image Rejection Ratio (IRR) of 58 dBc at 8 GHz, a power consumption of 50 mW, and an active area of 280 um × 180 um.