Suhas Kumar - Academia.edu (original) (raw)

Papers by Suhas Kumar

Scalable In-Memory Computing Architectures for Sparse Matrix Multiplication

2022 International Electron Devices Meeting (IEDM)

Activity-difference training of deep neural networks using memristor crossbars

Nature Electronics

Combinatorial Optimization in Hopfield Networks with Noise and Diagonal Perturbations

2022 IEEE International Symposium on Circuits and Systems (ISCAS)

Electro‐Thermal Characterization of Dynamical VO 2 Memristors via Local Activity Modeling

Advanced Materials

Physical Characterization of Current- and Temperature-Controlled Negative Differential Resistances in NbO 2

APS March Meeting Abstracts, Mar 1, 2017

Computing with device dynamics

Memristive Devices for Brain-Inspired Computing, 2020

Abstract Unlike the three fundamental circuit components, such as resistors, capacitors, and indu... more Abstract Unlike the three fundamental circuit components, such as resistors, capacitors, and inductors, which usually contain some linear function of current or voltage, memristors exhibit a truly and invariably nonlinear relationship between currents and voltages. This results in rich nonlinear dynamics embedded within individual components, which would otherwise require hundreds of transistors to emulate, for example, chaotic dynamics emerging from a single memristor. Meanwhile, in the world of mathematical modeling of the brain’s functioning, it has been shown that nearly all processes embody nonlinear behavior, from connections among neurons and synapses to the edge-of-chaos behavior giving rise to action potentials. In this chapter we discuss a few examples of how the rich nonlinear dynamics emerging from memristors can be used to construct brain-inspired (neuromorphic) computing systems, which remain a vastly unexplored topic that is gaining enormous attention of late.

arXiv (Cornell University), Nov 5, 2019

Articulation of a thought experiment in which the second law of thermodynamics appeared to be vio... more Articulation of a thought experiment in which the second law of thermodynamics appeared to be violated: "Maxwell's demon." Ludwig Boltzmann Statistical interpretation of entropy and the second law of thermodynamics. Josiah Gibbs Authoritative description of theories of thermodynamics, statistical mechanics and associated free energies and ensembles. Albert Einstein Theory of stochastic fluctuations displacing particles in a fluid: "Brownian Motion." John B. Johnson, Harry Nyquist Description of thermal fluctuation noise in electronic systems: "Johnson Noise." Lars Onsager Description of reciprocal relations among thermodynamic forces and fluxes in near equilibrium systems: "Onsager Relations." John von Neumann Developments of ergodic theory, quantum statistics, quantum entropy. Alan Turing Description of a minimalistic model of general computation: "Turing Machine." Claude Shannon Description of digital circuit design for Boolean operations. Claude Shannon Articulation of communications theory; foundations of information theory; connection of informational and physical concepts of entropy. John von Neumann Description of computing system architecture separating data and programs: the "Von Neumann Architecture."

Classical Adiabatic Annealing in Memristor Hopfield Neural Networks for Combinatorial Optimization

2020 International Conference on Rebooting Computing (ICRC), 2020

There is an intense search for supplements to digital computer processors to solve computationall... more There is an intense search for supplements to digital computer processors to solve computationally hard problems, such as gene sequencing. Quantum computing has gained popularity in this search, which exploits quantum tunneling to achieve adiabatic annealing. However, quantum annealing requires very low temperatures and precise control, which lead to unreasonably high costs. Here we show via simulations, alongside experimental instantiations, that computational advantages qualitatively similar to those gained by quantum annealing can be achieved at room temperature in classical systems by using a memristor Hopfield neural network to solve computationally hard problems.

Energy Efficient Computing R&D Roadmap Outline for Automated Vehicles

arXiv: Materials Science, 2020

Traditional electronic devices are well-known to improve in speed and energy-efficiency as their ... more Traditional electronic devices are well-known to improve in speed and energy-efficiency as their dimensions are reduced to the nanoscale. However, this scaling behavior remains unclear for nonlinear dynamical circuit elements, such as Mott neuron-like spiking oscillators, which are of interest for bio-inspired computing. Here we show that shrinking micrometer-sized VO2 oscillators to sub-100 nm effective sizes, achieved using a nanogap cut in a metallic carbon nanotube (CNT) electrode, does not guarantee faster spiking. However, an additional heat source such as Joule heating from the CNT, in combination with small size and heat capacity (defined by the narrow volume of VO2 whose insulator-metal transition is triggered by the CNT), can increase the spiking frequency by ~1000x due to an electro-thermal bifurcation in the nonlinear dynamics. These results demonstrate that nonlinear dynamical switches operate in a complex phase space which can be controlled by careful electro-thermal d...

Insights into the anomalous thermal properties of VO2 from synchrotron spectromicroscopy

Bulletin of the American Physical Society, 2019

ArXiv, 2019

We describe a hybrid analog-digital computing approach to solve important combinatorial optimizat... more We describe a hybrid analog-digital computing approach to solve important combinatorial optimization problems that leverages memristors (two-terminal nonvolatile memories). While previous memristor accelerators have had to minimize analog noise effects, we show that our optimization solver harnesses such noise as a computing resource. Here we describe a memristor-Hopfield Neural Network (mem-HNN) with massively parallel operations performed in a dense crossbar array. We provide experimental demonstrations solving NP-hard max-cut problems directly in analog crossbar arrays, and supplement this with experimentally-grounded simulations to explore scalability with problem size, providing the success probabilities, time and energy to solution, and interactions with intrinsic analog noise. Compared to fully digital approaches, and present-day quantum and optical accelerators, we forecast the mem-HNN to have over four orders of magnitude higher solution throughput per power consumption. Th...

Broadening the set of algorithms and use-cases for analog combinatorial optimization accelerators

Emerging Topics in Artificial Intelligence (ETAI) 2021, 2021

Recent experimental results show how classical accelerators based on analog computing can outperf... more Recent experimental results show how classical accelerators based on analog computing can outperform quantum annealing alternatives in benchmark tasks that require dense connection matrices. In Hewlett Packard Labs, we have been studying two alternatives: integrated coherent Ising machines and mem-HNNs (based on memristive crossbar arrays). An important challenge for commercial viability is that different industrial workloads typically benefit from the availability of a variety of optimization algorithms and require a broad range of template combinatorial optimization problems. In this talk, we will discuss our recent progress in going beyond Max-Cut, and we will propose a broader range of algorithms. This flexibility in algorithm choices and template problems is an important step forward to address the wide variety of enterprise-level use-cases such as airline scheduling, supply chain optimization, real-time bandwidth management, gene sequencing, etc.

IEEE Transactions on Circuits and Systems I: Regular Papers, 2021

We describe via simulation novel optimization algorithms for a Hopfield neural network constructe... more We describe via simulation novel optimization algorithms for a Hopfield neural network constructed using manufacturable three-terminal Silicon-Oxide-Nitride-Oxide-Silicon (SONOS) synaptic devices. We first present a computationally-light, memristor-based, highly accurate compact model for the SONOS. Using the compact model, we describe techniques of simulated annealing in Hopfield networks by exploiting imperfect problem definitions, current leakage, and the continuous tunability of the SONOS to enable transient chaotic group dynamics. We project improvements in energy consumption and latency for optimization relative to the best CPUs and GPUs by at least 4 orders of magnitude, and also exceeding the best projected memristor-based hardware; along with a 100-fold increase in error-resilient hardware size (i.e., problem size).

Future Computing Systems (FCS) to Support "Understanding" Capability

2019 IEEE International Conference on Rebooting Computing (ICRC), 2019

The massive explosion in data acquisition, processing, and archiving, accelerated by the end of M... more The massive explosion in data acquisition, processing, and archiving, accelerated by the end of Moore's Law, creates a challenge and an opportunity for a complete redesign of technology, devices, hardware architecture, software stack and AI stack to enable future computing systems with "understanding" capability. We propose a Future Computing System (FCS) based on a memory driven computing AI architecture, that leverages different types of next generation accelerators (e.g., Ising and Hopfield Machines), connected over an intelligent successor of the Gen-Z interconnect. On top of this architecture we propose a software stack and subsequently, an AI stack built on top of the software stack. While intelligence characteristics (learning, training, self-awareness, etc.) permeate all layers, we also separate AI-specific components into a separate layer for clear design. There are two aspects of AI in FCSs: a) AI embedded in the system to make the system better: better performing, more robust, self-healing, maintainable, repairable, and energy efficient. b) AI as the level of reasoning over the information contained within the system: the supervised and unsupervised techniques finding relationships over the data placed into the system. Developing the software and AI stack will require adapting to each redundant component. At least initially, specialization will be required. For this reason, starting with an interoperable, memory driven computing architecture and associated interconnect is essential for subsequent generalization. Our architecture is composable, i.e., it could be pursued in: a) its entirety, b) per-layer c) per component inside of the layer (e.g., only one of the accelerators, use cases, etc.); or d) exploring specific characteristics across the layers.

$Research paper thumbnail of Universal phase dynamics in VO 2 switches revealed by ultrafast operando diffraction$

Science, 2021

A transient metal Vanadium dioxide is known to have a coupled structural and electronic transitio... more A transient metal Vanadium dioxide is known to have a coupled structural and electronic transition that can be accessed through light, thermal, or electrical excitation. Ultrafast optical studies of this insulator-to-metal transition indicate that it is mediated by the formation of a transient metallic phase that retains the structure of the original insulating phase. Sood et al. show that a similar sequence occurs when the material is electrically excited with a series of voltage pulses. Using ultrafast electron diffraction, the researchers monitored the structure of a vanadium dioxide sample after excitation and found evidence of a metastable metallic phase that appears during the transition. Science , abc0652, this issue p. 352

Applied Physics Letters, 2019

Threshold switches, which typically exhibit an abrupt increase in current at an onset voltage, ha... more Threshold switches, which typically exhibit an abrupt increase in current at an onset voltage, have been used as selector devices to suppress leakage current in crosspoint arrays of two-terminal resistive switching memory devices. One of the most important metrics for selector devices is the leakage or low-voltage current, which limits the maximum achievable size of the crosspoint memory array. Here, we show that for self-heating-triggered threshold switches, there is an intrinsic lower limit to the leakage current resulting from the need to avoid an electric field-induced breakdown of the active material. We provide a quantitative theoretical estimate of this limit for NbO x threshold switches, one of the most widely studied selectors, and provide a plausible explanation for the experimentally observed leakage currents in NbO x. Our results provide some guidelines for achieving minimum leakage currents in threshold switches.

Applied Physics Reviews, 2020

Paper published as part of the special topic on Brain Inspired Electronics Note: This paper is pa... more Paper published as part of the special topic on Brain Inspired Electronics Note: This paper is part of the special collection on Brain Inspired Electronics.