Graph Neural Networks (GNNs) have exploded onto the machine learning scene in recent years owing to their capability to model and learn from graph-structured data. Such an ability has strong implications in a wide variety of fields whose data are inherently relational, for which conventional neural networks do not perform well. Indeed, as recent reviews can attest, research in the area of GNNs has grown rapidly and has lead to the development of a variety of GNN algorithm variants as well as to the exploration of ground-breaking applications in chemistry, neurology, electronics, or communication networks, among others. At the current stage research, however, the efficient processing of GNNs is still an open challenge for several reasons. Besides of their novelty, GNNs are hard to compute due to their dependence on the input graph, their combination of dense and very sparse operations, or the need to scale to huge graphs in some applications. In this context, this article aims to make two main contributions. On the one hand, a review of the field of GNNs is presented from the perspective of computing. This includes a brief tutorial on the GNN fundamentals, an overview of the evolution of the field in the last decade, and a summary of operations carried out in the multiple phases of different GNN algorithm variants. On the other hand, an in-depth analysis of current software and hardware acceleration schemes is provided, from which a hardware-software, graph-aware, and communication-centric vision for GNN accelerators is distilled.

This paper describes and validates for the first time the dynamic modelling of Liquid Crystal (LC)-based planar multi-resonant cells, as well as its use as bias signals synthesis tool to improve their reconfigurability time. The dynamic LC director equation is solved in the longitudinal direction through the finite elements method, which provides the z-and time-dependent inhomogeneous permittivity tensor used in an electromagnetic simulator to evaluate the cells behaviour. The proposed model has been experimentally validated using reflective cells for phase control (reflectarray) and measuring the transient phase, both in excitation and relaxation regimes. It is shown how a very reduced number of stratified layers are needed to model the material inhomogeneity, and that even an homogeneous effective tensor can be used in most of the cases, which allows a model simplification suitable for design procedures without losing accuracy. Consequently, a novel bias signal design tool is proposed to significantly reduce the transition times of LC cells, and hence, of electrically large antennas composed of them. These tools, similar to those used in optical displays, are experimentally validated for the first time at mm-and sub-mm wave frequencies in this work, obtaining an improvement of orders of magnitude.

The slow response time of planar Liquid Crystal (LC)-based phase-shift metasurface and Reconfigurable Intelligent Surfaces (RIS) cells is addressed in this paper by introducing a polymer network LC (PNLC) mixture suitable at mm-wave bands. Since the conventional effective isotropic model used in optical cells for describing the PNLC is not suitable in RF, an effective anisotropic and uniaxial model for such mixture is provided and experimentally validated at 100 GHz for the first time. In order to compare the temporal performance and tunability of the PNLC, transmissive and reflective cells, containing conventional LC and PNLC, have been manufactured and measured at optical and mm-wave frequencies. The temporal responses of PNLC are also compared for both RF and optical cells, obtaining relevant differences between their improvement factors, which are also discussed. Specifically, a 50 fold response time improvement is attained in cells designed to work at 100 GHz, although at the expense of a 3X tunability reduction. The model, which is robust to varying angle of incidence and cell dimensions, has been experimentally validated by designing and manufacturing a PNLC reflectarray cell of a different geometry. The cell shows reconfigurability times of 210ms, representing a significant improvement with respect to state-of-the-art response time in mm-wave cells, which are in the order of several seconds.

No abstract

<p>Liquid Crystal-based mm-wave spatially fed antennas with electronic reconfiguration are a promising solution at the higher frequencies required in next generation networks. However, one of the main drawbacks of the technology in these bands stems from the high reconfigurability times they present. Through this work, a relevant step towards overcoming the temporal problem by using Dual Frequency Liquid Crystals (DFLC) is presented. This paper details, for the first time, both the electromagnetic and temporal characterization of four commercially available DFLC mixtures in W-band, enabling their use in designing faster devices. To evaluate the experimental characterization, a reflectarray surface (made of 50x50 cells) specifically designed to achieve fast switching times with a sufficient phase range has been manufactured and measured. For this cell, a preliminary addressing technique based on overdriving has been used, exhibiting reconfigurability (rise and decay) times of 20 ms, one order of magnitude faster than the current state of the art of LC-based mm-wave planar devices. The measured results match the simulations, and reveal that a precisely designed biasing technique using overdrive must be used for DFLC-cells to achieve time reduction. Additionally, the benefits of this technology compared with other LC acceleration strategies in mm-wave are discussed.</p>

Recently, Graph Neural Networks (GNNs) have received a lot of interest because of their success in learning representations from graph structured data. However, GNNs exhibit different compute and memory characteristics compared to traditional Deep Neural Networks (DNNs). Graph convolutions require feature aggregations from neighboring nodes (known as the aggregation phase), which leads to highly irregular data accesses. GNNs also have a very regular compute phase that can be broken down to matrix multiplications (known as the combination phase). All recently proposed GNN accelerators utilize different dataflows and microarchitecture optimizations for these two phases. Different communication strategies between the two phases have been also used. However, as more custom GNN accelerators are proposed, the harder it is to qualitatively classify them and quantitatively contrast them. In this work, we present a taxonomy to describe several diverse dataflows for running GNN inference on accelerators. This provides a structured way to describe and compare the design-space of GNN accelerators.

Deep neural network (DNN) models continue to grow in size and complexity, demanding higher computational power to enable real-time inference. To efficiently deliver such computational demands, hardware accelerators are being developed and deployed across scales. This naturally requires an efficient scale-out mechanism for increasing compute density as required by the application. 2.5D integration over interposer has emerged as a promising solution, but as we show in this work, the limited interposer bandwidth and multiple hops in the Network-on-Package (NoP) can diminish the benefits of the approach. To cope with this challenge, we propose WIENNA, a wireless NoP-based 2.5D DNN accelerator. In WIENNA, the wireless NoP connects an array of DNN accelerator chiplets to the global buffer chiplet, providing highbandwidth multicasting capabilities. Here, we also identify the dataflow style that most efficienty exploits the wireless NoP's high-bandwidth multicasting capability on each layer. With modest area and power overheads, WIENNA achieves 2.2X-5.1X higher throughput and 38.2% lower energy than an interposer-based NoP design.

The main design principles in computer architecture have recently shifted from a monolithic scaling-driven approach to the development of heterogeneous architectures that tightly co-integrate multiple specialized processor and memory chiplets. In such data-hungry multi-chip architectures, current Networksin-Package (NiPs) may not be enough to cater to their heterogeneous and fast-changing communication demands. This position paper makes the case for wireless in-package nanonetworking as the enabler of efficient and versatile wired-wireless interconnect fabrics for massive heterogeneous processors. To that end, the use of graphene-based antennas and transceivers with unique frequency-beam reconfigurability in the terahertz band is proposed. The feasibility of such a nanonetworking vision and the main research challenges towards its realization are analyzed from the technological, communications, and computer architecture perspectives.