The goal of the NEUROKIT2E project is to create an open-source Deep Learning framework for edge and embedded AI built around an established European value chain. This framework, called AIDGE, supports a wide range of application areas that operate independently and serve a global user community. It provides easy and fast full-stack solutions from Neural Network design and optimization to AI application development all the way down to hardware implementations while enabling code generation for application-specific targets. This platform provides flexibility for academic users in the AI domain to explore and innovate while allowing them the possibility to prototype systems, ensuring their work aligns well with industrial needs. This paper presents the results and achievements of the first part of this three-year project, along with its roadmap and expected outcomes.

Timely identification of cardiac arrhythmia (abnormal heartbeats) is vital for early diagnosis of cardiovascular diseases. Wearable healthcare devices facilitate this process by recording heartbeats through electrocardiogram (ECG) signals and using AI-driven hardware to classify them into arrhythmia classes. Spiking neural networks (SNNs) are well-suited for such hardware as they consume low energy due to event-driven operation. However, their energy-efficiency and accuracy are constrained by encoding methods that translate real-valued ECG data into spikes. In this paper, we present an SNN-based ECG classification architecture featuring a new adaptive multi-threshold spike encoding scheme. This scheme adjusts encoding window and granularity based on the importance of ECG data samples, to capture essential information with fewer spikes. We develop a high-accuracy SNN model for such spike representation, by proposing a technique specifically tailored to our encoding. We design a hardware architecture for this model, which incorporates optimized layer post-processing for energy-efficient data-flow and employs fixed-point quantization for computational efficiency. Moreover, we integrate this architecture with our encoding scheme into a system-on-chip implementation using TSMC 40 nm technology. Our approach provides up to 5.1x energy-efficiency compared to state-of-the-art SNN-based ECG classifiers, with high accuracy.

Memristor-based Computation-In-Memory (CIM) has emerged as a compelling paradigm for designing energy-efficient neural network hardware. However, memristors suffer from conductance variation issue, which introduces computational errors in CIM hardware and leads to a degraded inference accuracy. In this paper, we present a hardware-aware quantization to mitigate the impact of conductance variation on CIM-based neural networks. We achieve this using the inherent characteristics of fixed-point arithmetic in CIM hardware. By tuning the bit-precision of weights, we align the conductance variation-induced errors with lower-order output bits. This reduces their numerical impact on the fixed-point output. We further decrease the residual errors by selectively discarding bits with low information and high error. This leads to error-free computations and a high inference accuracy. Our proposed methodology achieves 5.6× correct operations per unit energy compared to the conventional approach, while incurring very low hardware overheads.

Binary Neural Networks (BNNs) have demonstrated significant advantages in reducing computation and memory costs, all while maintaining acceptable accuracy on various image detection tasks. Thus, BNNs have the potential to support practical cognitive tasks on resource-constrained platforms, such as edge computing devices. To realize this, SRAM-based digital Computation-in-Memory (CIM) has gained growing attention as it overcomes the analog CIM architecture bottlenecks such as limited computing accuracy due to process variation, non-linearity, power and area-hungry Analog-to-Digital Converters (ADCs), etc. However, digital CIM architectures are highly dominated by power-hungry adder-trees, which can nullify the benefits of SRAM-based digital CIM. To address this issue, this paper proposes an adder free SRAM-based digital CIM, AFSRAM-CIM, for BNN acceleration. The proposed CIM architecture utilizes a multi-functional 10-T SRAM cell-based crossbar array and a new energy-efficient approach to perform the popcount operation. Simulation results using the MNIST dataset show that the proposed architecture maintains the state-of-the-art inference accuracy of 99.21% with only 11.86 fJ energy per operation. Moreover, AFSRAM-CIM achieves over 3× energy and ≈17× area savings when compared to the conventional digital CIM approaches.

Computation-in-memory (CIM) using memristors can facilitate data processing within the memory itself, leading to superior energy efficiency than conventional von-Neumann architecture. This makes CIM well-suited for data-intensive applications like neural networks. However, a large number of read operations can induce an undesired resistance change in the memristor, known as read-disturb. As memristor resistances represent the neural network weights in CIM hardware, read-disturb causes an unintended change in the network’s weights that leads to poor accuracy. In this paper, we propose a methodology for read-disturb detection and mitigation in CIM-based neural networks. We first analyze the key insights regarding the read-disturb phenomenon. We then introduce a mechanism to dynamically detect the occurrence of read-disturb in CIM-based neural networks. In response to such detections, we develop a method that adapts the sensing conditions of CIM hardware to provide error-free operation even in the presence of read-disturb. Simulation results show that our proposed methodology achieves up to 2× accuracy and up to 2× correct operations per unit energy compared to conventional CIM architectures.

Resistive Random Access Memories (RRAMs) are now undergoing commercialization, with substantial investment from many semiconductor companies. However, due to the immature manufacturing process, RRAMs are prone to exhibit new failure mechanisms and faults, which should be efficiently detected for high-volume production. Some of those faults are hard-to-detect, and require specific Design-for-Testability (DfT) circuit design. This paper proposes a DfT based on a parallel-reference write circuit that can detect all single-cell RRAM array faults: strong faults (directly causing logic errors) as well as weak faults (caused by parametric deviations). The scheme replaces the regular write driver, and enables the monitoring and comparison of the write current against multiple references during a single write operation. Hence, it serves as a DfT scheme and as a normal write circuit simultaneously. In addition, it enhances production testing speed and online fault detection, while keeping the area overhead low. Furthermore, the DfT is configurable for efficient diagnosis and yield learning. The results of the simulations performed do not only show that the DfT can detect single-cell conventional faults (due to interconnects and contacts) as well as unique RRAM faults (based on silicon data) that have been demonstrated to exist, but also that the DfT is robust to process variations.

Due to the immature manufacturing process, Resistive Random Access Memories (RRAMs) are prone to exhibit new failure mechanisms and faults, which should be efficiently detected for high-volume production. Those unique faults are hard to detect but require specific Design-for-Test (DfT) circuit design. This paper proposes a DfT based on a parallel-reference write circuit that can detect all RRAM array faults during diagnosis, production testing, and its application in the field.

Guaranteeing high-quality test solutions for Spin-Transfer Torque Magnetic RAM (STT-MRAM) is a must to speed up its high-volume production. A high test quality requires maximizing the fault coverage. Detecting permanent faults is relatively simple compared to intermittent faults; the latter are faults (caused by non-environmental conditions) that appear and disappear as a function of time, and are therefore hard to detect. Testing for such faults in STT-MRAMs is even worse considering the Magnetic Tunneling Junction inherent property ‘intrinsic switching stochasticity’, which results in inevitable random write errors. This paper presents a novel Design-for-Testability (DFT) scheme for detecting intermittent faults in STT-MRAMs; it is based on monitoring the write current. The strength of the write current is inversely correlated to the write error rate; when the write current is smaller than the specification, the device is considered faulty. A reduction in the write current can be caused by any defect in the write path of the memory (e.g., interconnects and contacts). Simulation results based on industrial design show that applying DFT yields a superior coverage of intermittent faults compared to functional test methods, such as march tests.

Resistive random access memory (RRAM) based computation-in-memory (CIM) architectures can meet the unprecedented energy efficiency requirements to execute AI algorithms directly on edge devices. However, the read-disturb problem associated with these architectures can lead to accumulated computational errors. To achieve the necessary level of computational accuracy, after a specific number of read cycles, these devices must undergo a reprogramming process which is a static approach and needs a large counter. This paper proposes a circuit-level RRAM read-disturb detection technique by monitoring real-time conductance drifts of RRAM devices, which initiate the reprogramming when actually it needs. Moreover, an analytic method is presented to determine the minimum conductance detection requirements, and our proposed read-disturb detection technique is tuned for the same to detect it dynamically. SPICE simulation result using TSMC 40 nm shows the correct functionality of our proposed detection technique.