-
Emerging Nonvolatile Memories for Machine Learning
Authors:
Adnan Mehonic,
Dovydas Joksas
Abstract:
Digital computers have been getting exponentially faster for decades, but huge challenges exist today. Transistor scaling, described by Moore's law, has been slowing down over the last few years, ending the era of fully predictable performance improvements. Furthermore, the data-centric computing demands fueled by machine learning applications are rapidly growing, and current computing systems --…
▽ More
Digital computers have been getting exponentially faster for decades, but huge challenges exist today. Transistor scaling, described by Moore's law, has been slowing down over the last few years, ending the era of fully predictable performance improvements. Furthermore, the data-centric computing demands fueled by machine learning applications are rapidly growing, and current computing systems -- even with the historical rate of improvements driven by Moore's law -- cannot keep up with these enormous computational demands. Some are turning to analogue in-memory computing as a solution, where specialised systems operating on physical principles accelerate specific tasks. We explore how emerging nonvolatile memories can be used to implement such systems tailored for machine learning. In particular, we discuss how memristive crossbar arrays can accelerate key linear algebra operations used in neural networks, what technological challenges remain, and how they can be overcome.
△ Less
Submitted 7 August, 2023;
originally announced August 2023.
-
Memristive, Spintronic, and 2D-Materials-Based Devices to Improve and Complement Computing Hardware
Authors:
Dovydas Joksas,
AbdulAziz AlMutairi,
Oscar Lee,
Murat Cubukcu,
Antonio Lombardo,
Hidekazu Kurebayashi,
Anthony J. Kenyon,
Adnan Mehonic
Abstract:
In a data-driven economy, virtually all industries benefit from advances in information technology -- powerful computing systems are critically important for rapid technological progress. However, this progress might be at risk of slowing down if we do not address the discrepancy between our current computing power demands and what the existing technologies can offer. Key limitations to improving…
▽ More
In a data-driven economy, virtually all industries benefit from advances in information technology -- powerful computing systems are critically important for rapid technological progress. However, this progress might be at risk of slowing down if we do not address the discrepancy between our current computing power demands and what the existing technologies can offer. Key limitations to improving energy efficiency are the excessive growth of data transfer costs associated with the von Neumann architecture and the fundamental limits of complementary metal-oxide-semiconductor (CMOS) technologies, such as transistors. In this perspective article, we discuss three technologies that will likely play an essential role in future computing systems: memristive electronics, spintronics, and electronics based on 2D materials. We present how these may transform conventional digital computers and contribute to the adoption of new paradigms, like neuromorphic computing.
△ Less
Submitted 1 July, 2022; v1 submitted 11 March, 2022;
originally announced March 2022.
-
Nonideality-Aware Training for Accurate and Robust Low-Power Memristive Neural Networks
Authors:
Dovydas Joksas,
Erwei Wang,
Nikolaos Barmpatsalos,
Wing H. Ng,
Anthony J. Kenyon,
George A. Constantinides,
Adnan Mehonic
Abstract:
Recent years have seen a rapid rise of artificial neural networks being employed in a number of cognitive tasks. The ever-increasing computing requirements of these structures have contributed to a desire for novel technologies and paradigms, including memristor-based hardware accelerators. Solutions based on memristive crossbars and analog data processing promise to improve the overall energy eff…
▽ More
Recent years have seen a rapid rise of artificial neural networks being employed in a number of cognitive tasks. The ever-increasing computing requirements of these structures have contributed to a desire for novel technologies and paradigms, including memristor-based hardware accelerators. Solutions based on memristive crossbars and analog data processing promise to improve the overall energy efficiency. However, memristor nonidealities can lead to the degradation of neural network accuracy, while the attempts to mitigate these negative effects often introduce design trade-offs, such as those between power and reliability. In this work, we design nonideality-aware training of memristor-based neural networks capable of dealing with the most common device nonidealities. We demonstrate the feasibility of using high-resistance devices that exhibit high $I$-$V$ nonlinearity -- by analyzing experimental data and employing nonideality-aware training, we estimate that the energy efficiency of memristive vector-matrix multipliers is improved by three orders of magnitude ($0.715\ \mathrm{TOPs}^{-1}\mathrm{W}^{-1}$ to $381\ \mathrm{TOPs}^{-1}\mathrm{W}^{-1}$) while maintaining similar accuracy. We show that associating the parameters of neural networks with individual memristors allows to bias these devices towards less conductive states through regularization of the corresponding optimization problem, while modifying the validation procedure leads to more reliable estimates of performance. We demonstrate the universality and robustness of our approach when dealing with a wide range of nonidealities.
△ Less
Submitted 5 May, 2022; v1 submitted 13 December, 2021;
originally announced December 2021.
-
Committee machines -- a universal method to deal with non-idealities in memristor-based neural networks
Authors:
D. Joksas,
P. Freitas,
Z. Chai,
W. H. Ng,
M. Buckwell,
C. Li,
W. D. Zhang,
Q. Xia,
A. J. Kenyon,
A. Mehonic
Abstract:
Artificial neural networks are notoriously power- and time-consuming when implemented on conventional von Neumann computing systems. Consequently, recent years have seen an emergence of research in machine learning hardware that strives to bring memory and computing closer together. A popular approach is to realise artificial neural networks in hardware by implementing their synaptic weights using…
▽ More
Artificial neural networks are notoriously power- and time-consuming when implemented on conventional von Neumann computing systems. Consequently, recent years have seen an emergence of research in machine learning hardware that strives to bring memory and computing closer together. A popular approach is to realise artificial neural networks in hardware by implementing their synaptic weights using memristive devices. However, various device- and system-level non-idealities usually prevent these physical implementations from achieving high inference accuracy. We suggest applying a well-known concept in computer science -- committee machines -- in the context of memristor-based neural networks. Using simulations and experimental data from three different types of memristive devices, we show that committee machines employing ensemble averaging can successfully increase inference accuracy in physically implemented neural networks that suffer from faulty devices, device-to-device variability, random telegraph noise and line resistance. Importantly, we demonstrate that the accuracy can be improved even without increasing the total number of memristors.
△ Less
Submitted 27 August, 2020; v1 submitted 14 September, 2019;
originally announced September 2019.
