Currently, Machine Learning (ML) is becoming ubiquitous in everyday life. Deep Learning (DL) is already present in many applications ranging from computer vision for medicine to autonomous driving of modern cars as well as other sectors in security, healthcare, and finance. However, to achieve impressive performance, these algorithms employ very deep networks, requiring a significant computational power, both during the training and inference time. A single inference of a DL model may require billions of multiply-and-accumulated operations, making the DL extremely compute-and energy-hungry. In a scenario where several sophisticated algorithms need to be executed with limited energy and low latency, the need for cost-effective hardware platforms capable of implementing energy-efficient DL execution arises. This paper first introduces the key properties of two brain-inspired models like Deep Neural Network (DNN), and Spiking Neural Network (SNN), and then analyzes techniques to produce efficient and high-performance designs. This work summarizes and compares the works for four leading platforms for the execution of algorithms such as CPU, GPU, FPGA and ASIC describing the main solutions of the state-of-the-art, giving much prominence to the last two solutions since they offer greater design flexibility and bear the potential of high energy-efficiency, especially for the inference process. In addition to hardware solutions, this paper discusses some of the important security issues that these DNN and SNN models may have during their execution, and offers a comprehensive section on benchmarking, explaining how to assess the quality of different networks and hardware systems designed for them.

Deep Neural Networks (DNNs) are nowadays a common practice in most of the Artificial Intelligence (AI) applications. Their ability to go beyond human precision has made these networks a milestone in the history of AI. However, while on the one hand they present cutting edge performance, on the other hand they require enormous computing power. For this reason, numerous optimization techniques at the hardware and software level, and specialized architectures, have been developed to process these models with high performance and power/energy efficiency without affecting their accuracy. In the past, multiple surveys have been reported to provide an overview of different architectures and optimization techniques for efficient execution of Deep Learning (DL) algorithms. This work aims at providing an up-to-date survey, especially covering the prominent works from the last 3 years of the hardware architectures research for DNNs. In this paper, the reader will first understand what a hardware accelerator is, and what are its main components, followed by the latest techniques in the field of dataflow, reconfigurability, variable bit-width, and sparsity.

In today’s world, ruled by a great amount of data and mobile devices, cloud-based systems are spreading all over. Such phenomenon increases the number of connected devices, broadcast bandwidth, and information exchange. These fine-grained interconnected systems, which enable the Internet connectivity for an extremely large number of facilities (far beyond the current number of devices) go by the name of Internet of Things (IoT). In this scenario, mobile devices have an operating time which is proportional to the battery capacity, the number of operations performed per cycle and the amount of exchanged data. Since the transmission of data to a central cloud represents a very energy-hungry operation, new computational paradigms have been implemented. The computation is not completely performed in the cloud, distributing the power load among the nodes of the system, and data are compressed to reduce the transmitted power requirements. In the edge-computing paradigm, part of the computational power is moved toward data collection sources, and, only after a first elaboration, collected data are sent to the central cloud server. Indeed, the “edge” term refers to the extremities of systems represented by IoT devices. This survey paper presents the hardware architectures of typical IoT devices and sums up many of the low power techniques which make them appealing for a large scale of applications. An overview of the newest research topics is discussed, besides a final example of a complete functioning system, embedding all the introduced features.

Transformers' compute- intensive operations pose enormous challenges for their deployment in resource- constrained EdgeAI / tiny ML devices. As an established neural network compression technique, quantization reduces the hardware computational and memory resources. In particular, fixed-point quantization is desirable to ease the computations using lightweight blocks, like adders and multipliers, of the underlying hardware. However, deploying fully-quantized Transformers on existing general-purpose hardware, generic AI accelerators, or specialized architectures for Transformers with floating-point units might be infeasible and/or inefficient. Towards this, we propose SwiftTron, an efficient specialized hardware accelerator designed for Quantized Transformers. SwiftTron supports the execution of different types of Transformers' operations (like Attention, Softmax, GELU, and Layer Normalization) and accounts for diverse scaling factors to perform correct computations. We synthesize the complete SwiftTron architecture in a 65 nm CMOS technology with the ASIC design flow. Our Accelerator executes the RoBERTa-base model in 1.83 ns, while consuming 33.64 mW power, and occupying an area of 273 mm ² • To ease the reproducibility, the RTL of our SwiftTron architecture is released at https://github.com/albertomarchisio/SwiftTron.

In recent years, the development of Advanced Driver-Assistance Systems (ADASs) is driving the need for more reliable and precise on-vehicle sensing. Radar and lidar are crucial in this framework, since they allow sensing of vehicle’s surroundings. In such a scenario, it is necessary to master these sensing systems, and knowing their similarities and differences is important. Due to ADAS’s intrinsic real-time performance requirements, it is almost mandatory to be aware of the processing algorithms required by radar and lidar to understand what can be optimized and what actions can be taken to approach the real-time requirement. This review aims to present state-of-the-art radar and lidar technology, mainly focusing on modulation schemes and imaging systems, highlighting their weaknesses and strengths. Then, an overview of the sensor data processing algorithms is provided, with some considerations on what type of algorithms can be accelerated in hardware, pointing to some implementations from the literature. In conclusion, the basic concepts of sensor fusion are presented, and a comparison between radar and lidar is performed.

The widespread adoption of data-centric algorithms, particularly Artificial Intelligence (AI) and Machine Learning (ML), has exposed the limitations of centralized processing infrastructures, driving a shift towards edge computing. This necessitates stringent constraints on energy efficiency, which traditional von Neumann architectures struggle to meet. The Compute-In-Memory (CIM) paradigm has emerged as a better candidate due to its efficient exploitation of the available memory bandwidth. However, existing CIM solutions require a high implementation effort and lack flexibility from a software integration standpoint. This work proposes a novel, software-friendly, general-purpose, and low-integration-effort Near-Memory Computing (NMC) approach, paving the way for the adoption of CIMbased systems in the next generation of edge computing nodes. Two architectural variants, NM-Caesar and NM-Carus, are proposed and characterized to target different trade-offs in area efficiency, performance, and flexibility, covering a wide range of embedded microcontrollers. Post-layout simulations show up to 28.0× and 53.9× lower execution time and 25.0× and 35.6× higher energy efficiency at system level, respectively, compared to the execution of the same tasks on a state-of-the-art RISC-V CPU (RV32IMC). NM-Carus achieves a peak energy efficiency of 306.7GOPS/W in 8-bit matrix multiplications, surpassing recent state-of-the-art in- and near-memory circuits.