In-Memory Computing (IMC) has emerged as one of the most promising architectures to efficiently compute artificial intelligence tasks on hardware, particularly Deep Neural Networks (DNNs). IMC can make use of analog computation principles alongside emerging Non-Volatile Memory (eNVM) technologies, potentially offering several orders of magnitude increased energy efficiency compared to generic processing units. Yet, the use of analog circuitry, potentially integrated with emerging technologies post-processed on top of silicon wafers, increases the susceptibility of hardware to a large spectrum of variations, for instance manufacturing, noise or temperature sensitivity. Hence, this susceptibility can hamper the large-scale deployment of IMC circuits into the market. To tackle the reliability of analog resistive-based IMC circuits regarding temperature variations, this paper presents TRIM, a thermal on-chip auto-compensation method aimed at fully calibrating first-order temperature effects. TRIM is designed to maintain the computational accuracy of IMC cores in DNN applications over a wide temperature range, while being highly scalable and adaptable. In essence, the temperature compensation is realized through a Complementary-To-Absolute-Temperature (CTAT) voltage reference integrated inside a voltage regulator and applied at the zero reference node of a Multiplying Digital-to-Analog Converter (MDAC), eliminating the need for external circuits or look-up tables. The proposed methodology is demonstrated on a proof-of-concept 65 nm CMOS resistive IMC column. Measurement results showcase that the proof-of-concept auto-compensation system significantly enhances inference and Multiply-And-Accumulate (MAC) operation accuracy of any first-order resistive crossbar column, achieving inference accuracy recovery of 100% over a temperature range of -20 ∘C to 60 ∘C and a 91.3 in MAC operation accuracy, with an area overhead of 2% and power overhead of <0.02%.

Probabilistic models (PMs) have become an alternative to complement or replace deep learning in applications where transparency and trustworthiness are crucial. As PMs compute explicit high-resolution probabilities, ensuring numerical stability legitimates the need for logarithmic (log) computing. As exact log computation on hardware is typically costly, existing hardware accelerators stick to high-resolution linear computation with,e.g., floating point (FP). From the perspective of efficient execution on edge devices, using such generic linear hardware for log operations is prone to underflow and ill-suited for operations such as log addition. Hence, the log-domain computing of PMs requires new hardware solutions, combining numerical stability and energy-efficient execution. Inspired by the Log-Sum-Exp (LSE) function used in existing PM software tools transferring data between log and linear domains to compute log additions, this work proposes an LSE Processing Element (LSE-PE). LSE-PE allows for efficient log computation, through an innovative double approximation for log addition, while ensuring numerical stability with an error compensation method using a compact error correction Look-Up Table (CLUT). Hardware synthesis results using a 16nm technology show that the proposed 24-bit LSE-PE hardware consumes 46% area and 32% power of 32-bit floating point, using only 16 LUT entries with 10 bits in each entry. Moreover, our experiments on various PM benchmarks show that LSE-PE prevents underflow even for large models, which exist in all other 32-bit number systems, with less than 0.2% accuracy loss. We also demonstrate an outlier detection task for uncertainty estimation of image classification models using the LSE-PE, for a fraction of the main model’s computing cost (0.06 to 20% of representative DNN architectures for MNIST).

In-memory computing (IMC) accelerators have become a pivotal architecture for enhancing AI algorithm computations, particularly critical for embedding deep neural networks (DNNs) in edge devices. The efficiency of these systems is paramount, yet IMC cores are prone to fluctuations due to process, temperature, and voltage variations, which can detrimentally impact DNN accuracy. This research introduces an innovative Built-In Self-Calibration (BISC) methodology, specifically designed to compensate for temperature-induced variations in mixed-signal IMC cores. The methodology enables real-time, on-chip adjustment of DNN weights during computation within the IMC core without modifying the computation path. The proposed approach, implemented on a silicon prototype, not only maintained DNN computation accuracy under substantial temperature variations but also fully compensated for almost 90% of the offset caused by these variations, without introducing any non-idealities.

Probabilistic circuits (PCs) offer a promising avenue to perform embedded reasoning under uncertainty. They support efficient and exact computation of various probabilistic inference tasks by design. Hence, hardware-efficient computation of PCs is highly interesting for edge computing applications. As computations in PCs are based on arithmetic with probability values, they are typically performed in the log domain to avoid underflow. Unfortunately, performing the log operation on hardware is costly. Hence, prior work has focused on computations in the linear domain, resulting in high resolution and energy requirements. This work proposes the first dedicated approximate computing framework for PCs that allows for low-resolution logarithm computations. We leverage Addition As Int, resulting in linear PC computation with simple hardware elements. Further, we provide a theoretical approximation error analysis and present an error compensation mechanism. Empirically, our method obtains up to 357{\texttimes} and 649{\texttimes} energy reduction on custom hardware for evidence and MAP queries respectively with little or no computational error.

Deep neural network (DNN) accelerators are increasingly integrated into sensing applications, such as wearables and sensor networks, to provide advanced in-sensor processing capabilities. Given wearables’ strict size and power requirements, minimizing the area and energy consumption of DNN accelerators is a critical concern. In that regard, computing DNN models in the time domain is a promising architecture, taking advantage of both technology scaling friendliness and efficiency. Yet, time-domain accelerators are typically not fully digital, limiting the full benefits of time-domain computation. In this work, we propose an all-digital time-domain accelerator with a small size and low energy consumption to target precision in-sensor processing like human activity recognition (HAR). The proposed accelerator features a simple and efficient architecture without dependencies on analog nonidealities such as leakage and charge errors. An eight-neuron layer (core computation layer) is implemented in 22-nm FD-SOI technology. The layer occupies$70 \times \,70\,\mu $m while supporting multibit inputs (8-bit) and weights (8-bit) with signed accumulation up to 18 bits. The power dissipation of the computation layer is 576$\mu $W at 0.72-V supply and 500-MHz clock frequency achieving an average area efficiency of 24.74 GOPS/mm2 (up to 544.22 GOPS/mm2), an average energy efficiency of 0.21 TOPS/W (up to 4.63 TOPS/W), and a normalized energy efficiency of 13.46 1b-TOPS/W (up to 296.30 1b-TOPS/W).

In-Memory Computing (IMC) architectures promise increased energy-efficiency for embedded artificial intelligence. Many IMC circuits rely on analog computation, which is more sensitive to process and temperature variations than digital. Thus, maintaining a suitable computation accuracy may require process and temperature compensation. Focusing on resistive-based IMC architectures, we propose an ultra-low power circuit to compensate for the temperature and process-based non-linearities of resistive computing elements. The proposed circuit, implemented in 65 nm CMOS can provide a temperature coefficient between 10 and 1938 ppm/°C for a wide temperature range (-40°C to 80°C) and output current range (few pA up to 600 nA) at 1.2 V operating voltage. Used in a resistive IMC array, the variation of output currents from each multiply-accumulate (MAC) operation can be reduced by up to 84% to maintain computation accuracy across process and temperature variations.

In-Memory Computing (IMC) accelerators based on resistive crossbars are emerging as a promising pathway toward improved energy efficiency in artificial neural networks. While significant research efforts are directed toward designing advanced resistive memory devices, the nonidealities associated with practical device implementation are often overlooked. Existing solutions typically compensate for these nonidealities during off-chip training, introducing additional complexities and failing to account for random errors such as noise, device failures, and cycle-to-cycle variability. To tackle this challenge, this work proposes a self-calibrated activation neuron topology that offers a fully online non-linearity compensation for IMC accelerators. The neuron merges multiply-accumulate operations with Rectified Linear Unit (ReLU) activation function in the analog domain for increased efficiency. The self-calibration is integrated into the data conversion process to minimize overheads and be fully online. The proposed activation neuron is designed and simulated using 22 nm FDSOI CMOS technology. The design demonstrates robustness across a wide temperature range (-40°C to 80°C) and under various process corners, with a maximum accuracy loss of 1 LSB for an 8-bit activation accuracy.

With tightening automotive IC production test requirements, test escape rates need to decrease down to the 10 PPB level. To achieve this for mixed-signal ICs, advanced multivariate statistical techniques are needed, as the defects in the test escapes become increasingly more difficult to detect. Therefore, this paper proposes applying a cascade of advanced statistical techniques to identify measurements that can be used as predictors to flag future potential failures at test time with minimal misclassification of good devices. The approach uses measurement data from the ATE wafer probe tests and is also able to identify the likely location of the defect using only these measurements. The cascade has four steps: 1) remove bias and spatial patterns within the data, 2) divide the different tests into relevant groups, 3) reduce the dimensionality of each group, and 4) perform multiple regression to find the predictor values and use these values to compute an outlier score for each chip under test. As there is a risk of overfitting the outlier score, the number of predictors used is kept to a minimum. The effectiveness of the proposed methodology is demonstrated using test data from an industrial production chip with eight field-return cases. Predictors have been found that retroactively allowed the identification of these chips, with an average of 5% false classification of good devices, i.e. devices not returned from the field. In addition, the selected predictors corresponded to where the defects are located according to failure analysis of the field returns.

Safety-critical and mission-critical systems, such as airplanes or (semi-)autonomous cars, are relying on an ever-increasing number of embedded integrated circuits. Consequently, there is a need for complete defect coverage during the testing of these circuits to guarantee their functionality in the field. In this context, reducing the escape rate of defects during production testing is crucial, and significant progress has been made to this end. However, production testing using automatic test equipment is subject to various measurement parasitic variations, which may have a negative impact on the testing procedure and therefore limit the final defect coverage. To tackle this issue, this article proposes an improved test flow targeting increased analog defect coverage, both at the system and block levels, by analyzing and improving the coverage of typical functional and structural tests under these measurement variations. To illustrate the flow, the technique of inserting a pseudo-random signal at available circuit nodes and applying machine learning techniques to its response is presented. A DC-DC converter, derived from an industrial product, is used as a case study to validate the flow. In short, results show that system-level tests for the converter suffer strongly from the measurement variations and are limited to just under 80% coverage, even when applying the proposed test flow. Block-level testing, however, can achieve only 70% fault coverage without improvements but is able to consistently achieve 98% of fault coverage at a cost of at most 2% yield loss with the proposed machine learning–based boosting technique.

The numerous variations that affect analog and RF circuits are becoming a limiting factor in the design of these circuits in deeply scaled CMOS technologies. An emerging idea to counteract these effects is to let the circuit compensate for these variations itself, referred to as self-healing. Over the last decade, a wide variety of off- and on-chip techniques for compensating these variations have been researched. This paper targets to give an overview of the state-of-the-art, and organize the proposed techniques in a common taxonomy. This allows to determine remaining open issues and research challenges. In particular, the SotA lacks efficient solutions for fully-integrated, short time-scale self-adaptation. The paper ends by giving an outlook towards promising research directions to enable such self-adaptation in mWatt power budgets for Internet of things applications, focusing on embedded machine-learning techniques.

Over the past few years, several self-calibration methodologies have proven their efficiency to calibrate analog and radio-frequency circuits against process variations. Specifically, statistical techniques based on machine-learning have been proposed to recover yield loss and even enhance circuit performances. In addition, these techniques enable to calibrate circuits after a single performance test, i.e. in one-shot. However, towards fully-integrated calibration techniques, the inference part of the machine learning algorithm needs to be performed as energy-efficiently as possible to reduce calibration cost to a minimum. Following the path of resource-efficient machine learning, this work explores an alternative to state-of-the-art Neural Network based statistical techniques. Specifically, we investigate the opportunities of using Bayesian Networks for resource-efficient on-chip statistical calibration of analog/RF circuits. Results will show that several improvements can be achieved using Bayesian Networks: (a) provide a comprehensive calibration framework with explicit relationships between parameters (b) demonstrate similar prediction accuracies that neural networks (c) optimize across several performance parameters with a single network and in a single query and (d) enable a more energy-efficient hardware implementation. The proposed self-calibration algorithm is applied to a low-noise amplifier fabricated with IBM's 130nm CMOS process, leading to a significant reduction in the number of operations required to obtain the best tuning knob setting.

The Phoenix1project aims to develop a new approach to explore unknown environments, based on multiple measurement campaigns carried out by extremely tiny devices, called agents, that gather data through multiple sensors. These low power and low resource agents are configured specifically for each measurement campaign to achieve the exploration goal in the smallest number of iterations. Thus, the main design challenge is to build agents as much reconfigurable as possible. This paper introduces the Phoenix project in more details, and presents first developments in the agent design.

We propose a post-fabrication calibration technique for RF circuits that is performed during production testing with minimum extra cost. Calibration is enabled by equipping the circuit with tuning knobs and sensors. Optimal tuning knob identification is achieved in one-shot based on a single test step that involves measuring the sensor outputs once. For this purpose, we rely on variation-aware sensors which provide measurements that remain invariant under tuning knob changes. As an auxiliary benefit, the variation-aware sensors are non-intrusive and totally transparent to the circuit. The technique is demonstrated on a 65nm RF power amplifier.

In the context of mission-critical, safety-critical, and remote-controlled applications, it is required to equip systems with self-adapting capabilities. Adaptation is required in post-manufacturing to correct yield loss and achieve zero defective parts-per-million as well as during normal operation to account for different application scenarios and for varying environmental conditions. A self-adaptive system must be capable of providing the required high performances after manufacturing and throughout its normal operation regardless the application scenario wherein it is deployed and despite the varying environmental conditions. In this paper, we describe a generic post-manufacturing self-adaptation technique for RF circuits as well as concurrent self-adaptation techniques for a safety-critical medical sensor for glaucoma diagnosis and for a NFC system which is very sensitive to the environment in which it operates.