Low-energy Computing Research Articles

Multi-level, forming and filament free, bulk switching trilayer RRAM for neuromorphic computing at the edge

CMOS-RRAM integration holds great promise for low energy and high throughput neuromorphic computing. However, most RRAM technologies relying on filamentary switching suffer from variations and noise, leading to computational accuracy loss, increased energy consumption, and overhead by expensive program and verify schemes. We developed a filament-free, bulk switching RRAM technology to address these challenges. We systematically engineered a trilayer metal-oxide stack and investigated the switching characteristics of RRAM with varying thicknesses and oxygen vacancy distributions to achieve reliable bulk switching without any filament formation. We demonstrated bulk switching at megaohm regime with high current nonlinearity, up to 100 levels without compliance current. We developed a neuromorphic compute-in-memory platform and showcased edge computing by implementing a spiking neural network for an autonomous navigation/racing task. Our work addresses challenges posed by existing RRAM technologies and paves the way for neuromorphic computing at the edge under strict size, weight, and power constraints.

Booth Encoding-Based Energy Efficient Multipliers for Deep Learning Systems

Artificial intelligence on edge is a growing research field. In this paper, we propose a novel re-encoding scheme for reducing the size of the weights of deep neural networks (DNNs). The proposed re-encoding scheme exploits the Booth encoding scheme and the power-of-two (PO2) quantization to allow for very low energy computations during the inference of the neural networks with minimal loss in classification accuracy. We demonstrate the advantages of the proposed re-encoding scheme by computing a convolutional neural network (CNN) and a linear neural network on the proposed Extended Exact Multiplier and the proposed PO2 Multiplier. Our proposed PO2 quantization and re-encoding method reduce the model size for the CNN by 30.77% and the model size of the linear neural network by 49.86%. Furthermore, our multipliers reduce the inference energy for CNN by 50.6% and for the linear neural network by 90.1%. The PO2 Multiplier is proposed for the sensor-end computation of the linear neural network with a 77.32% reduction in the area relative to an exact Booth multiplier and it reduces the inference energy consumption of the linear neural network by 93.2% when compared to the unmodified exact multiplier. Our proposed scheme can be used to improve the energy consumption during inference for most Booth multipliers with minor modifications to the re-encoding signal arrangements. We also demonstrate that the proposed re-encoding scheme paired with the proposed multipliers outperforms all the existing designs in terms of resource utilization with a minimal impact on the inference accuracy of the neural networks.

A DQN-Based Frame Aggregation and Task Offloading Approach for Edge-Enabled IoMT

The rapid expansion of wearable medical devices and health data of Internet of Medical Things (IoMT) poses new challenges to the high Quality of Service (QoS) of intelligent health care in the foreseeable 6G era. Healthcare applications and services require ultra reliable, ultra low delay and energy consumption data communication and computing. Wireless Body Area Network (WBAN) and Mobile Edge Computing (MEC) technologies empowered IoMT to deal with huge data sensing, processing and transmission in high QoS. However, traditional frame aggregation schemes in WBAN generate too much control frames during data transmission, which leads to high delay and energy consumption and is not flexible enough. In this paper, a Deep Q-learning Network (DQN) based Frame Aggregation and Task Offloading Approach (DQN-FATOA) is proposed. Firstly, different service data were divided into queues with similar QoS requirements. Then, the length of the frame aggregation was selected dynamically by the aggregation node according to the delay, energy consumption, and throughput by DQN. Finally, the number of tasks offloaded was selected due to the current state. Compared with the traditional scheme, the simulation results show that the proposed DQN-FATOA has effectively reduced delay and energy consumption, and improved the throughput and overall utilization of WBAN.

A Self-Attentional ResNet-LightGBM Model for IoT-Enabled Voice Liveness Detection

Voice user interface (VUI) brings high efficiency and convenience for the applications of Internet of Things (IoTs), meanwhile, it can also cause increasingly serious security issues. The word-level voice liveness detection is proved to be the promising solution to thwart spoofing attacks. However, The complex acoustic feature, diversified attacks, and different interaction distance can severely affects the improvement of detection accuracy. To alleviate this issue, we develop a novel pop noise based word-level voice liveness detection framework. Firstly, a new voice frame selection method is proposed for determining optimal frames, including short time Fourier transform, low frequency average energy computation and sequencing. Then, the acoustic features of the selected frames are calculated by Gammatone frequency cepstral coefficient (GFCC). Finally, based on these features, a newly-built joint voice detector, fusing the self-attentional residual network (ResNet) and light gradient boosting machine (LightGBM), can achieve accurate voice classification. On the popular voice spoofing attack datasets, experimental results show that our proposal significantly outperforming the baseline and the state of the arts models, and it is gender dependent. Moreover, our proposal has good generalization ability for far-field replay voice attack, speech synthesis and voice conversion attacks, and partial fake voice attack. Finally, its effectiveness is verified by the ablation study.

Epitaxial Integration of Dirac Semimetals with Si(001)

Topological semimetals contain novel combinations of properties that make them useful in a variety of applications, including optoelectronics, spintronics and low energy computing, and catalysis. Although they have been grown with high quality as bulk single crystals, incorporation with semiconductor substrates will ultimately be required to maximize their technological reach. Here, epitaxial growth of the Dirac semimetal Cd3As2 on Si(001) is demonstrated through two routes. First, Cd3As2(112) epilayers are grown on Si(001) via an intermediate CdTe(111) buffer layer. Second, Cd3As2(112) is grown directly on Si(001). This work sets the foundation for integration of novel semimetal materials with existing CMOS technology.

Batteryless IoT module for sensing and signaling failures of passive power accessories

This paper presents a design of a batteryless Internet of Things (IoT) module for upgrading power lines accessories such as surge arresters and insulators with responsive capabilities. The module is able to power itself from the failure current and to signal the nature and consequences of the failed accessory to a remote dispatch. It comprises an energizing circuit, a signal shaping circuitry, a custom-developed continuity indicator, and a low power communication microprocessor with programmable settings. The paper provides an elaboration of the constraints under which such a device must operate, simulation results on its feasibility, analytical calculation of optimal placement of certain components in 3D and measurement results to support the underlying theory of its operation. The research is an original contribution to the field of customized low energy computing and cyber-physical systems incorporating optimized microprocessor solutions applied to the IoT concept.

Symbiotic Organisms Search Optimization to Predict Optimal Thread Count for Multi-threaded Applications

Multicore systems have emerged as a cost-effective option for the growing demands for high-performance, low-energy computing. Thread management has long been a source of concern for developers, as overheads associated with it reduce the overall throughput of the multicore processor systems. One of the most complex problems with multicore processors is determining the optimal number of threads for the execution of multithreaded programs. To address this issue, this paper proposes a novel solution based on a modified symbiotic organism search (MSOS) algorithm which is a bio-inspired algorithm used for optimization in various engineering domains. This technique uses mutualism, commensalism and parasitism behaviours seen in organisms for searching the optimal solutions in the available search space. The algorithm is simulated on the NVIDIA DGX Intel-Xeon E5-2698-v4 server with PARSEC 3.0 benchmark suit. The results show that keeping the thread count equal to the number of processors available in the system is not necessarily the best strategy to get maximum speedup when running multithreaded programs. It was also observed that when programs are run with the optimal thread count, the execution time is substantially decreased, resulting in energy savings due to the use of fewer processors than are available in the system.

Function Virtualization Can Play a Great Role in Blockchain Consensus

Bitcoin introduced a cryptocurrency as a form of public ledger consequently that turned into a most popular security technology, Blockchain. Its integrated mining technology lies the key security mechanism. The system allows forming a pool mining group to solve a particular job and share their revenues to their CPU usage while one of them successfully mines a block. To mine a block, a cryptographic puzzle should be solved, which requires significant compute resources that cause huge energy consumption. On the other hand, recent statistics show that low computational energy-restricted Internet of Things (IoT) devices are increasing exponentially. Although it has low energy and limited computation power, it is large in quantity when it is integrated. So we focus on a stochastic geometry theory, which resolves the issue of block mining computation via utilizing multiple mobile IoT devices, given that these IoT devices are Computation Capable Nodes (CCNs). To further normalize this issue, we propose an efficient mathematical solution that uses smart coordination of Virtual Network Functions (VNFs) for IoT devices to enable their CPU usage efficiently. At the same time, the work and credit point distribution policy is smartly handled through virtual pool mining. The proposal renders Network Function Virtualization technology to configure VNF, and Service Function Chain technology is utilized to enable the network flow of such VNFs. New algorithms are presented to solve multiple issues like node discovery, computation offloading, and work credit point distribution. Our goal is to minimize energy consumption within the given time constraint. Implementation results show that although virtual functions for block mining require extensive computations in IoT devices, dividing computation work into small fractions called tasks embedded with VNF, and offloading them to nearby CCNs, tend to minimize the cost and energy consumption of individual shared miners. The overall mining process is proved efficient and faster.

Efficient Point-Process Modeling of Spiking Neurons for Neuroprosthesis.

Neuroprosthesis refers to implantable medical devices which can replace injured biological functions in the brain. One of the core problems in neuroprosthesis study is to construct a neural signal transformation model from one cortical area to another. Since the brain encodes and transmits information in spike trains, spiking neural network (SNN) can be an ideal choice for neuroprosthesis modeling. This paper proposes a spiking neuron point-process model (SNPM), which receives spike times as input, and is capable of modeling nonlinear interactions between cortical areas. The proposed SNPM can be implemented on neuromorphic chips for low-energy computing, thus has potential for clinical applications. Experiments show that SNPM can accurately reconstruct functional relationships from PMd (dorsal premotor cortex) to M1 (primary motor cortex) areas.

Materials opportunities for low-energy computing

This issue focuses on the materials opportunities for reducing energy consumption in the different stacks of the compute hardware—devices, memory, interconnects, and microelectronics packaging. The thermodynamics of computing to understand the minimum “compute energy limits” that can be aspired is briefly reviewed. One of the most long-standing limits that remain largely ignored from a technology standpoint, the Landauer limit, is discussed. The relationship between thermodynamics and information, starting with Maxwell’s paradox, is introduced.

Characterisation and Modelling of Moisture Gradients in Polyamide 6.

Polyamide 6 (PA6) is able to absorb water from the surrounding air and bond to it by forming hydrogen bonds between the carbonamide groups of its molecular chains. Diffusion processes cause locally different water concentrations in the (component) cross-section during the sorption process, resulting in locally different mechanical properties due to the water-induced plasticisation effect. However, the water content of PA6 is usually specified as an integral value, so no information about a local water distribution within a component is provided. This paper shows a method to characterise moisture distributions within PA6 samples using low-energy computer tomography (CT) techniques and comparing the reconstructed results with a developed finite elements (FE) modelling method based on Fick’s diffusion laws with concentration-dependent diffusion coefficients. For this purpose, the ageing of the samples at two different water bath temperatures as well as at different integral water contents are considered. The results obtained by CT reconstruction and FE modelling are in very good agreement, so that the concentration distributions by water sorption of PA6 calculated by FEM can be regarded as validated.

Principles of low dissipation computing from a stochastic circuit model

We introduce a thermodynamically consistent, minimal stochastic model for complementary logic gates built with field-effect transistors. We characterize the performance of such gates with tools from information theory and study the interplay between accuracy, speed, and dissipation of computations. With a few universal building blocks, such as the NOT and NAND gates, we are able to model arbitrary combinatorial and sequential logic circuits, which are modularized to implement computing tasks. We find generically that high accuracy can be achieved provided sufficient energy consumption and time to perform the computation. However, for low-energy computing, accuracy and speed are coupled in a way that depends on the device architecture and task. Our work bridges the gap between the engineering of low dissipation digital devices and theoretical developments in stochastic thermodynamics, and provides a platform to study design principles for low dissipation digital devices.

Optical configuration of an N ∶ 2N reversible decoder using a LiNbO3-based Mach-Zehnder interferometer.

These days when integrated circuit (IC) designers are facing an uphill task in limiting energy/heat dissipation, reversible computing is emerging as a potential candidate with vast application in fields like nanotechnology, quantum-dot cellular automata, and low power IC. Optical reversible logics have turned up to offer high speed and low energy computations with almost no loss of input information when a certain (arithmetic or logical) operation is performed. This paper explores an optical implementation of an optimized Fredkin gate that is employed to design an $ N:2^N $ reversible decoder. The optical designs have been carried out using the electro-optic effect of a lithium niobate ($ {{\rm LiNbO}_3}$)-based Mach-Zehnder interferometer under the beam propagation method (BPM) with Optiwave's OptiBPM tool. The mathematical model of output power of these designs is also performed along with its validation in MATLAB.

Long-range spin-wave propagation in transversely magnetized nano-scaled conduits

Magnonics attracts increasing attention in the view of low-energy computation technologies based on spin waves. Recently, spin-wave propagation in longitudinally magnetized nano-scaled spin-wave conduits was demonstrated, proving the fundamental feasibility of magnonics at the sub-100 nm scale. Transversely magnetized nano-conduits, which are of great interest in this regard as they offer a large group velocity and a potentially chirality-based protected transport of energy, have not yet been investigated due to their complex internal magnetic field distribution. Here, we present a study of propagating spin waves in a transversely magnetized nanoscopic yttrium iron garnet conduit of 50 nm width. Space and time-resolved microfocused Brillouin-light-scattering spectroscopy is employed to measure the spin-wave group velocity and decay length. A long-range spin-wave propagation is observed with a decay length of up to (8.0 ± 1.5) μm and a large spin-wave lifetime of up to (44.7 ± 9.1) ns. The results are supported with micromagnetic simulations, revealing a broad single-mode frequency range and the absence of a mode localized to the edges. Furthermore, a frequency nonreciprocity for counter-propagating spin waves is observed in the simulations and the experiment, caused by the trapezoidal cross section of the structure. The revealed long-distance spin-wave propagation on the nano-scale is particularly interesting for an application in spin-wave devices, allowing for long-distance transport of information in magnonic circuits and low-energy device architectures.

Systems Science: Digital Transformation andDeveloping Business Model toward Smart Farms’platform

This paper describes efforts by National Authority for Remote Sensing and Space Sciences (NARSS) to help the Egyptian government to manage and monitor the national projects. We successfully developed a geospatial data sharing portal (NARSSGeoPortal) as part of the government need to build national Decision Support System (DSS). We were able to solve the software development issues as well as the satellite imagery sourcing issues, but the main challenge remains around how to collect complete and correct data from the public about their private businesses nationwide. The most challenging is how to engage the public and encourage the business owners who are the main sources of data to provide the government Geoportal with data about their businesses. It is also challenging to engage the scientists and experts from government research centers into the data sharing Geoportal. Furthermore, it is a challenge to integrate the government research centers with the public businesses’ daily operation. The data sharing Geoportal is built for all national projects and government authorities, however, in this paper we focus on the Agriculture authorities and farming businesses where the challenge is how to collect correct and complete data per acre about the seeds, fertilizers, water, pest control and all other farm related data that the satellite imagery does not provide. The goal is to integrate the farms into unified national monitoring, and control system while developing advanced smart farms with the use of Internet of Things (IoT). The proposed collaboration agriculture platform fills the gap between two groups. The first group includes the government authorities, financial institutions, and research centers. The second group includes farmers, supply chain, and agriculture engineers. The platform show how employment can be generated by transforming the national ecosystem. The paper also fills a major gap in industry as well as in academia by providing the first Bluetooth Low Energy computer aided design tool that will facilitate testing, designing, deploying, managing and debugging of real Bluetooth Low Energy networks.

A training algorithm for networks of high-variability reservoirs

Physical reservoir computing approaches have gained increased attention in recent years due to their potential for low-energy high-performance computing. Despite recent successes, there are bounds to what one can achieve simply by making physical reservoirs larger. Therefore, we argue that a switch from single-reservoir computing to multi-reservoir and even deep physical reservoir computing is desirable. Given that error backpropagation cannot be used directly to train a large class of multi-reservoir systems, we propose an alternative framework that combines the power of backpropagation with the speed and simplicity of classic training algorithms. In this work we report our findings on a conducted experiment to evaluate the general feasibility of our approach. We train a network of 3 Echo State Networks to perform the well-known NARMA-10 task, where we use intermediate targets derived through backpropagation. Our results indicate that our proposed method is well-suited to train multi-reservoir systems in an efficient way.

Sparse coding with a somato-dendritic rule

Cortical neurons are silent most of the time: sparse activity enables low-energy computation in the brain, and promises to do the same in neuromorphic hardware. Beyond power efficiency, sparse codes have favourable properties for associative learning, as they can store more information than local codes but are easier to read out than dense codes. Auto-encoders with a sparse constraint can learn sparse codes, and so can single-layer networks that combine recurrent inhibition with unsupervised Hebbian learning. But the latter usually require fast homeostatic plasticity, which could lead to catastrophic forgetting in embodied agents that learn continuously. Here we set out to explore whether plasticity at recurrent inhibitory synapses could take up that role instead, regulating both the population sparseness and the firing rates of individual neurons. We put the idea to the test in a network that employs compartmentalised inputs to solve the task: rate-based dendritic compartments integrate the feedforward input, while spiking integrate-and-fire somas compete through recurrent inhibition. A somato-dendritic learning rule allows somatic inhibition to modulate nonlinear Hebbian learning in the dendrites. Trained on MNIST digits and natural images, the network discovers independent components that form a sparse encoding of the input and support linear decoding. These findings confirm that intrinsic homeostatic plasticity is not strictly required for regulating sparseness: inhibitory synaptic plasticity can have the same effect. Our work illustrates the usefulness of compartmentalised inputs, and makes the case for moving beyond point neuron models in artificial spiking neural networks.

Materials opportunities and challenges for low-energy computing: Thermal management and interconnects

Abstract

Active Nanophotonics [Scanning the Issue

Being able to manipulate and control light flows at small scales holds the promise to open groundbreaking opportunities for a variety of technologies. Consider, for instance, the challenges currently faced in the world of computing: as data rates and processing demands increase worldwide at an exponential rate, we are facing unsustainable increases in energy consumption associated with data centers and streaming providers. To address these challenges, optical computing and communications offer an interesting alternative to electronic-based systems. Using light for these purposes, however, is hindered by the fact that photons are not easily squeezed to volumes beyond the diffraction limit, i.e., below the wavelength scale, which would be required both to enable lowenergy computation at sufficiently high speeds and to match the degree of integration density available in electronic systems. Today, the field of optoelectronics, which combines highdensity electronic devices to process the data and low-energy data transport enabled by light, is growing at a very fast pace.

Experimental investigation of silicon and silicon nitride platforms for phase-change photonic in-memory computing

Advances in artificial intelligence have greatly increased demand for data-intensive computing. Integrated photonics is a promising approach to meet this demand in big-data processing due to its potential for wide bandwidth, high speed, low latency, and low-energy computing. Photonic computing using phase-change materials combines the benefits of integrated photonics and co-located data storage, which of late has evolved rapidly as an emerging area of interest. In spite of rapid advances of demonstrations in this field on both silicon and silicon nitride platforms, a clear pathway towards choosing between the two has been lacking. In this paper, we systematically evaluate and compare computation performance of phase-change photonics on a silicon platform and a silicon nitride platform. Our experimental results show that while silicon platforms are superior to silicon nitride in terms of potential for integration, modulation speed, and device footprint, they require trade-offs in terms of energy efficiency. We then successfully demonstrate single-pulse modulation using phase-change optical memory on silicon photonic waveguides and demonstrate efficient programming, memory retention, and readout of > 4 bits of data per cell. Our approach paves the way for in-memory computing on the silicon photonic platform.