Although offering great potential for energy-efficient edge-AI, memristor-based CIM accelerators are severely hindered by IR drop induced errors. To tackle this, we propose a low-cost mitigation technique by first quantifying the impact of IR drop on the accuracy. Then, a mitigation strategy is developed to compensate for IR drop-induced inference accuracy reduction by combining an optimized mapping scheme with a fine-tuned calibration of the ADC. Results show the proposed solution can effectively mitigate IR drop with a negligible overhead.

The Computing-In-Memory (CIM) paradigm offers a promising solution to the memory-wall bottleneck that limits conventional Von Neumann architectures. By performing data processing at the same physical location where the data are stored, CIM-based architectures minimize costly data movement and drastically improve energy efficiency. When implemented with Ferroelectric Field Effect Transistors (FeFETs), additional advantages from the non-volatility, fast switching, and low operating voltage of FeFETs are added. However, the widespread adoption of FeFETs is limited by their poor endurance, which is overcome by a Back End of the Line (BEoL) integration of FeFET-2, where a ferroelectric capacitor (FeCAP) is wired to the gate of a CMOS transistor providing high endurance compatible with low-power edge applications. These properties enable dense, low-power, and high-speed matrix operations essential for AI workloads. As a result, FeFET-2-based CIM accelerators offer a promising solution for energy-efficient, high-performance AI at the edge. The Ferro4EdgeAI project aims to develop an ultra low-power, scalable edge accelerator for AI, targeting a significant gain in energy efficiency with respect to state-of-the-art AI hardware accelerators. To attain this, our project focuses on innovation all along the value chain from materials, physic concepts, device architecture, integration technologies, and accelerators in a holistic design space exploration approach.

Computing-in-Memory (CIM) architectures using memristive crossbar arrays enable energy-efficient AI acceleration. Analog non-idealities, such as IR drop and nonlinearity, impose design constraints that existing simulators cannot capture and thus explore effectively. Current approaches sacrifice either modeling accuracy or simulation speed, preventing systematic design space exploration. In this paper we propose X-Sim, a crossbar simulator that resolves this trade-off through a modular architecture. Our approach decouples device physics from circuit analysis using a fixed-point scheme, avoiding expensive Jacobian computations while preserving device fidelity. X-Sim delivers SPICE-level accuracy (< 1% error) with up to 200× speedup over physics-based simulators. This enables quick and systematic design space exploration across thousands of configurations, guiding reliable system design. X-Sim will be released as open source.

Advancements in Artificial Intelligence (AI) and Internet-of-Things (IoT) have increased demand for edge AI, but deployment on traditional AI accelerators, like GPUs and TPUs, using von Neumann architecture, suffer from inefficiencies due to separate memory and compute units. Computation-in-Memory (CIM), utilizing non-volatile memristor devices to leverage analog computing principles and perform in-place computations, holds great potential in improving computational efficiency by eliminating frequent data movement. However, standard implementation of CIM faces several challenges, primarily high power consumption and subsequently induced nonlinearity, debating its viability for edge devices. In this paper, we propose C3CIM, a novel memristor-based CIM micro-architecture, featuring a new bit-cell and array design, targeting efficient implementation of Neural Networks (NN). Our architecture uses a constant current source to perform Multiply-and-Accumulate (MAC) operations with a very low computation current (10 to 100 nA), thereby significantly enhancing power efficiency. We adapted C3CIM for Spiking Neural Networks (SNN) and developed a prototype using TSMC 40nm CMOS node for on-silicon validation. Furthermore, our micro-architecture was benchmarked using two SNN models based on N-MNIST and IBM-Gesture datasets, for comparison against current state-of-the-art (SOTA). Results show up to 35x reduction in power along with 6.7x saving in energy compared to SOTA, demonstrating promising potential of this work for edge AI applications.

International audience

We propose an on-line testing methodology for neuromorphic hardware supporting spiking neural networks. Testing aims at detecting in real-time abnormal operation due to hardware-level faults, as well as screening of outlier or corner inputs that are prone to misprediction. Testing is enabled by two on-chip classifiers that prognosticate, based on a low-dimensional set of features extracted with spike counting, whether the network will make a correct prediction. The system of classifiers is capable of evaluating the confidence of the decision, and when the confidence is judged low a replay operation helps to resolve the ambiguity. The testing methodology is demonstrated by fully embedding it in a custom FPGA-based neuromorphic hardware platform. It operates in the background being totally non-intrusive to the network operation, while offering a zero-latency test decision for the vast majority of inferences.

We address the problem of testing artificial intelligence (AI) hardware accelerators implementing spiking neural networks (SNNs). We define a metric to quickly rank available samples for training and testing based on their fault detection capability. The metric measures the interclass spike count difference of a sample for the fault-free design. In particular, each sample is assigned a score equal to the spike count difference between the first two top classes. The hypothesis is that samples with small scores achieve high fault coverage because they are prone to misclassification, i.e., a small perturbation in the network due to a fault will result in these samples being misclassified with high probability. We show that the proposed metric correlates with the per-sample fault coverage and that retaining a set of high-ranked samples in the order of ten achieves near-perfect fault coverage for critical faults that affect the SNN accuracy. The proposed test generation approach is demonstrated on two SNNs modeled in Python and on actual neuromorphic hardware. We discuss fault modeling and perform an analysis to reduce the fault space so as to speed up test generation time.

Despite the parallelism and sparsity in neural network models, their transfer into hardware unavoidably makes them susceptible to hardware-level faults. Hardware-level faults can occur either during manufacturing, such as physical defects and process-induced variations, or in the field due to environmental factors and aging. The performance under fault scenarios needs to be assessed so as to develop cost-effective fault-tolerance schemes. In this work, we assess the resilience characteristics of a hardware accelerator for Spiking Neural Networks (SNNs) designed in VHDL and implemented on an FPGA. The fault injection experiments pinpoint the parts of the design that need to be protected against faults, as well as the parts that are inherently fault-tolerant.

The error-resiliency of Artificial Intelligence (AI) hardware accelerators is a major concern, especially when they are deployed in mission-critical and safety-critical applications. In this paper, we propose a neuron fault tolerance strategy for Spiking Neural Networks (SNNs). It is optimized for low area and power overhead by leveraging observations made from a large-scale fault injection experiment that pinpoints the critical fault types and locations. We describe the fault modeling approach, the fault injection framework, the results of the fault injection experiment, the fault-tolerance strategy, and the fault-tolerant SNN architecture. The idea is demonstrated on two SNNs that we designed for two SNN-oriented datasets, namely the N-MNIST and IBM's DVS128 gesture datasets.

The deployment of Artificial Intelligence (AI) hardware accelerators in a variety of applications, including safety-critical ones, requires assessing their inherent reliability to hardware-level faults and developing cost-effective fault tolerance techniques. This entails performing large-scale fault simulation experiments. However, transistor-level fault simulation is prohibitive and fault simulation should be carried out at a higher abstraction level. In this work, we focus on spiking neural networks (SNNs), and we follow a bottom-up approach starting from transistor-level simulations for developing a neuron behavioral-level fault model that can be readily employed for performing behavioral-level fault simulation of deep SNNs.