We propose a fully integrated common-source amplifier based analog artificial neural network (ANN). The performance of the proposed ANN with a custom non-linear activation function is demonstrated on the breast cancer classification task. A hardware-software co-design methodology is adopted to ensure good matching between the software AI model and hardware prototype. A 65 nm prototype of the proposed ANN is fabricated and characterized. The prototype ANN achieves 97% classification accuracy when operating from a 1.1 V supply with an energy consumption of 160 fJ/classification. The prototype consumes 50 W power and occupies 0.003 mm die area.
21 February 2019 in IEEE Transactions on Circuits and Systems I: Regular Papers
A continuous-time second-order ΔΣ analog-to-digital converter (ADC) is presented in this paper. The proposed ADC is based on a novel architecture and uses current starved ring oscillators as integrators to achieve the second-order noise shaping. The proposed architecture does not require excess loop delay compensation or nonlinearity calibration. Static element mismatch in the multi-bit current digital-to-analog converter is high-pass shaped by intrinsic data-weighted averaging. Detailed analysis and insights on the various trade-offs involved in the design of the proposed ADC are presented in this paper. A prototype ADC is implemented in 65-nm CMOS and achieves 64.2-dB SNDR at a bandwidth of 2.5 MHz and a Schreier FoM of 158.2 dB. The ADC operates at 205 MHz and consumes 1 mW of power. The measured power supply rejection ratio is 56.2 dB.
1 November 2020 in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
This work presents a unified weak physical unclonable function (PUF) and a true random number generator (TRNG) based on the current-steering digital-to-analog converter (DAC) and ring voltage-controlled oscillator (VCO). Entropy source for the weak PUF is the mismatch between NMOS and PMOS transistors in a cascode current DAC as well as the mismatch between VCO quantizers, while entropy source for the TRNG is thermal noise in the DAC and VCO and clock jitter. Instead of using spatial entropy sources for the PUF, i.e., multiple unit PUF elements, the proposed architecture utilizes temporal entropy source by capturing the output of unit PUF element over multiple cycles, which reduces area significantly. A unified PUF/TRNG prototype is fabricated in 65-nm CMOS and consumes 0.36 pJ/bit at a throughput of 100 Mb/s. The PUF has a measured intra-HD of 0.0906 and inter-HD of 0.4859, while the raw TRNG bitstream has an entropy of 0.9991 and passes all the NIST statistical randomness tests.
This work presents a true random number generator (TRNG) based on a stochastic delta-sigma modulator (DSM) with ring oscillator (RO) integrators. The TRNG leverages noise and jitter in the ROs as entropy source. A multi-bit non-return-to-zero (NRZ) digital-to-analog converter (DAC) in the feedback path ensures DSM noise dominates input swing seen by the front-end integrators. Hence, the DSM can concurrently operate as an ADC and TRNG. A 65nm prototype achieves Shannon entropy >0.999998 with lower-bound min-entropy >0.995 at 52Mbps throughput while passing all NIST tests across multiple chips and voltage/temperature corners without needing calibration. The combined power consumption of ADC and TRNG is 0.36mW which is the lowest among state-of-the-art RO TRNGs. Low power consumption, high entropy, and concurrent operation of the TRNG as an ADC are principal contributions of this work.
17 August 2020 in IEEE Solid-State Circuits Letters
This work presents the first on-chip bio-inspired reservoir computer (RC) prototype implemented in 65nm CMOS. The RC comprises 50 time-multiplexed neurons, and each neuron embeds a strong non-linearity in a feedback loop. The RC applies a non-linear transformation to the input and projects it to high-dimensional space, thus allowing linear separation by a simple logistic regression (LR) layer implemented off-chip. We demonstrate real-time stress detection from electrocardiogram (ECG) signals using the RC. The RC achieves 93% classification accuracy which is 6% better than state-of-the-art digital classifiers. Operating at 40kHz, the prototype consumes 27.5nJ/classification which is 7× lower than state-of-the-art ECG processors performing similar complexity classification task.
28 January 2021 in IEEE Transactions on Circuits and Systems I: Regular Papers
This paper presents fully-integrated analog neural network classifier architecture for low resolution image classification that eliminates memory access. We design custom activation functions using single-stage common-source amplifiers, and apply a hardware-software co-design methodology to incorporate knowledge of the custom activation functions into the training phase to achieve high accuracy. Performing all computations entirely in the analog domain eliminates energy cost associated with memory access and data movement. We demonstrate our classifier on multinomial classification task of recognizing downsampled handwritten digits from MNIST dataset. Fabricated in 65nm CMOS process, the measured energy consumption for down-sampled MNIST dataset is 173pJ/classification, which is 3× better than state-of-the-art. The prototype IC achieves mean classification accuracy of 81.3% even after down-sampling the original MNIST images by 96% from 28 × 28 pixels to 5 × 5 pixels.