Search

Recent advances in artificial intelligence have prompted the search for enhanced algorithms and hardware to support the deployment of machine learning at the edge. More specifically, in the context of the Internet of Things (IoT), vision chips must be able to fulfill tasks of low to medium complexity, such as feature extraction or region-of-interest (RoI) detection, with a sub-mW power budget imposed by the use of small batteries or energy harvesting. Mixed-signal vision chips relying on in- or near-sensor processing have emerged as an interesting candidate, thanks to their favorable tradeoff between energy efficiency (EE) and computational accuracy compared to digital systems for these specific tasks. In this paper, we introduce a mixed-signal convolutional imager system-on-chip (SoC) codenamed MANTIS, featuring a unique combination of large 16$\times$16 4b-weighted filters, operation at multiple scales, and double sampling, well suited to the requirements of medium-complexity tasks. The main contributions are (i) circuits called DS3 units combining delta-reset sampling, image downsampling, and voltage downshifting, and (ii) charge-domain multiply-and-accumulate (MAC) operations based on switched-capacitor amplifiers and charge sharing in the capacitive DAC of the successive-approximation ADCs. MANTIS achieves peak EEs normalized to 1b operations of 4.6 and 84.1 TOPS/W at the accelerator and SoC levels, while computing feature maps with a root mean square error ranging from 3 to 11.3$\%$. It also demonstrates a face RoI detection with a false negative rate of 11.5$\%$, while discarding 81.3$\%$ of image patches and reducing the data transmitted off chip by 13$\times$ compared to the raw image., Comment: 14 pages, 23 figures

In recent years, the development of the Internet of Things (IoT) has prompted the search for nA-range current references that are simultaneously constrained to a small area and robust to process, voltage and temperature variations. Yet, such references have remained elusive, as existing architectures fail to reach a low temperature coefficient (TC) while minimizing silicon area. In this work, we propose a nA-range constant-with-temperature (CWT) peaking current reference, in which a resistor is biased by the threshold voltage difference between two transistors in weak inversion. This bias voltage is lower than in conventional architectures to cut down the silicon area occupied by the resistor and is obtained by forward body biasing one of the two transistors with an ultra-low-power voltage reference so as to reduce its threshold voltage. In addition, the proposed reference includes a circuit to suppress the leakage of parasitic diodes at high temperature, and two simple trimming mechanisms for the reference current and its TC. As the proposed design relies on the body effect, it has been validated in both 0.11-$\mu$m bulk and 22-nm fully-depleted silicon-on-insulator, to demonstrate feasibility across different technology types. In post-layout simulation, the 0.11-$\mu$m design generates a 5-nA current with a 65-ppm/{\deg}C TC and a 2.84-%/V line sensitivity (LS), while in measurement, the 22-nm design achieves a 1.5-nA current with an 89-ppm/{\deg}C TC and a 0.51-%/V LS. As a result of the low resistor bias voltage, the proposed references occupy a silicon area of 0.00954 mm$^2$ in 0.11 $\mu$m (resp. 0.00214 mm$^2$ in 22 nm) at least 1.8$\times$ (resp. 8.2$\times$) smaller than fabricated nA-range CWT references, but with a TC improved by 6.1$\times$ (resp. 4.4$\times$)., Comment: 14 pages, 28 figures

Internet-of-Things (IoT) applications require nW-power current references that are robust to process, voltage and temperature (PVT) variations, to maintain the performance of IoT sensor nodes in a wide range of operating conditions. However, nA-range current references are rarely area-efficient due to the use of large gate-leakage transistors or resistors, which occupy a significant silicon area at this current level. In this paper, we introduce a nA-range constant-with-temperature (CWT) current reference, relying on a self-cascode MOSFET (SCM) biased by a four-transistor ultra-low-power voltage reference through a single-transistor buffer. The proposed reference includes a temperature coefficient (TC) calibration mechanism to maintain performance across process corners. In addition, as the proposed design relies on the body effect, it has been fabricated and measured in 0.11-$\mu$m bulk and 22-nm fully-depleted silicon-on-insulator (FD-SOI) to demonstrate feasibility in both technology types. On the one hand, the 0.11-$\mu$m design consumes a power of 16.8 nW at 1.2 V and achieves a 2.3-nA current with a line sensitivity (LS) of 2.23 %/V at 25{\deg}C and a TC of 176 ppm/{\deg}C at 1.2 V from -40 to 85{\deg}C. On the other hand, the 22-nm design consumes a power of 16.3 nW at 1.5 V and achieves a 2.5-nA current with a 1.53-%/V LS at 25{\deg}C and an 82-ppm/{\deg}C TC at 1.5 V from -40 to 85{\deg}C. Thanks to their simple architecture, the proposed references achieve a silicon area of 0.0106 mm$^2$ in 0.11 $\mu$m and 0.0026 mm$^2$ in 22 nm without compromising other figures of merit, and are thus competitive with state-of-the-art CWT references operating in the same current range., Comment: 14 pages, 26 figures

In many applications, the ability of current references to cope with process, voltage, and temperature (PVT) variations is critical to maintaining system-level performance. However, temperature-independent current references operating in the nA range are rarely area-efficient due to the use of large resistors which occupy a significant silicon area at this current level. In this article, we introduce a nA-range constant-with-temperature (CWT) current reference relying on a self-cascode MOSFET (SCM), biased by a proportional-to-absolute-temperature (PTAT) voltage with a CWT offset. On the one hand, the proposed reference has been simulated post-layout in 65-nm bulk. This design consumes 5.4 nW at 0.7 V and achieves a 1.1-nA current with a line sensitivity (LS) of 0.69 %/V and a temperature coefficient (TC) of 213 ppm/$^\circ$C. On the other hand, the proposed reference has been simulated and fabricated in 22-nm fully depleted silicon-on-insulator (FDSOI). This second design requires additional features to mitigate the impact of parasitic diode leakage at high temperature. In measurement, it consumes 5.8 nW at 0.9 V and achieves a 0.9-nA current with a 0.39-%/V LS and a 565-ppm/$^\circ$C TC. As a result of using an SCM, the proposed references occupy a silicon area of 0.0021 mm$^2$ in 65 nm (respectively, 0.0132 mm$^2$ in 22 nm) at least 25$\times$ (respectively, 4$\times$) smaller than state-of-the-art CWT references operating in the same current range., Comment: 13 pages, 25 figures

The robustness of current and voltage references to process, voltage and temperature (PVT) variations is paramount to the operation of integrated circuits in real-world conditions. However, while recent voltage references can meet most of these requirements with a handful of transistors, current references remain rather complex, requiring significant design time and silicon area. In this paper, we present a family of simple current references consisting of a two-transistor (2T) ultra-low-power voltage reference, buffered onto a voltage-to-current converter by a single transistor. Two topologies are fabricated in a 0.18-$\mu$m partially-depleted silicon-on-insulator (SOI) technology and measured over 10 dies. First, a 7T nA-range proportional-to-absolute-temperature (PTAT) reference intended for constant-$g_m$ biasing of subthreshold operational amplifiers demonstrates a 0.096-nA current with a line sensitivity (LS) of 1.48 %/V, a temperature coefficient (TC) of 0.75 %/$^\circ$C, and a variability $(\sigma/\mu)$ of 1.66 %. Then, two 4T+1R $\mu$A-range constant-with-temperature (CWT) references with (resp. without) TC calibration exhibit a 1.09-$\mu$A (resp. 0.99-$\mu$A) current with a 0.21-%/V (resp. 0.20-%/V) LS, a 38-ppm/$^\circ$C (resp. 290-ppm/$^\circ$C) TC, and a 0.87-% (resp. 0.65-%) $(\sigma/\mu)$. In addition, portability to common scaled CMOS technologies, such as 65-nm bulk and 28-nm fully-depleted SOI, is discussed and validated through post-layout simulations., Comment: 14 pages, 25 figures

While the backpropagation of error algorithm enables deep neural network training, it implies (i) bidirectional synaptic weight transport and (ii) update locking until the forward and backward passes are completed. Not only do these constraints preclude biological plausibility, but they also hinder the development of low-cost adaptive smart sensors at the edge, as they severely constrain memory accesses and entail buffering overhead. In this work, we show that the one-hot-encoded labels provided in supervised classification problems, denoted as targets, can be viewed as a proxy for the error sign. Therefore, their fixed random projections enable a layerwise feedforward training of the hidden layers, thus solving the weight transport and update locking problems while relaxing the computational and memory requirements. Based on these observations, we propose the direct random target projection (DRTP) algorithm and demonstrate that it provides a tradeoff between accuracy and computational cost that is suitable for adaptive edge computing devices., Comment: This document is the paper as accepted for publication in the Frontiers in Neuroscience journal, the fully-edited paper is available at https://www.frontiersin.org/articles/10.3389/fnins.2021.629892

Shifting computing architectures from von Neumann to event-based spiking neural networks (SNNs) uncovers new opportunities for low-power processing of sensory data in applications such as vision or sensorimotor control. Exploring roads toward cognitive SNNs requires the design of compact, low-power and versatile experimentation platforms with the key requirement of online learning in order to adapt and learn new features in uncontrolled environments. However, embedding online learning in SNNs is currently hindered by high incurred complexity and area overheads. In this work, we present ODIN, a 0.086-mm$^2$ 64k-synapse 256-neuron online-learning digital spiking neuromorphic processor in 28nm FDSOI CMOS achieving a minimum energy per synaptic operation (SOP) of 12.7pJ. It leverages an efficient implementation of the spike-driven synaptic plasticity (SDSP) learning rule for high-density embedded online learning with only 0.68$\mu$m$^2$ per 4-bit synapse. Neurons can be independently configured as a standard leaky integrate-and-fire (LIF) model or as a custom phenomenological model that emulates the 20 Izhikevich behaviors found in biological spiking neurons. Using a single presentation of 6k 16$\times$16 MNIST training images to a single-layer fully-connected 10-neuron network with on-chip SDSP-based learning, ODIN achieves a classification accuracy of 84.5% while consuming only 15nJ/inference at 0.55V using rank order coding. ODIN thus enables further developments toward cognitive neuromorphic devices for low-power, adaptive and low-cost processing., Comment: (C) 2018 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses. This document is the paper as accepted for publication in the IEEE Transactions on Biomedical Circuits and Systems journal, the fully-edited paper is available at https://ieeexplore.ieee.org/document/8528875