RKRajkumar Kubendran

Rajkumar Kubendran

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Rajkumar Kubendran has a total of 77 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Online Charge Balancing With Active Impedance Monitoring in Programmable Stimulator to Extend Electrode Lifetime

    by Kangni Liu, Guangzong Chen, Kevin Woeppel, X. Tracy Cui, Rajkumar Kubendran

    Neurostimulation therapies are often applied as an alternative method to pharmaceutical treatment for pain relief.This paper demonstrates a programmable stimulator for analgesic nerve stimulation through invasive electrodes. It provides two modes of operation: constant-current biphasic and capacitor-coupled biphasic. The stimulator is integrated on a 50 x 30 mm Printed Circuit Board (PCB) and can generate up to ±1.6 mA current pulses in steps of 6.3 μA. A custom STM32 microcontroller library is used to tune the pulse duration range from 1 μs to 10 s and stimulation frequency range between 0.1 Hz to 100 kHz. An Active Impedance Monitoring (AIM) unit in the PCB measures the frequency response from 1 Hz to 10 kHz with 8 samples per decade every 30 minutes. The electrode resistance, tissue resistance and RC time constant of the electrode-tissue system could be calculated from the frequency response curve, generated by an input square wave stimulation current at different frequencies. In different measurements with discrete RC circuit components, the PCB could calculate the impedance values with >90% precision. A novel charge-balancing algorithm implemented on the STM32 microcontroller could match the positive and negative charge within a 5% error. Through online adjustment of stimulation parameters using active impedance monitoring to achieve charge balancing, the proposed programmable stimulation system extends the electrode lifetime for more than 288 million continuous pulses in vitro studies. The in vitro studies clearly demonstrated the efficacy of charge balanced stimulation that enable the electrodes to last around 3 times longer with negligible degradation.

  • Top Cited Papers

    Neuromorphic computing at scale

    by Dhireesha Kudithipudi, Catherine Schuman, Craig M. Vineyard, Tej Pandit, Cory Merkel, Rajkumar Kubendran, James B. Aimone, Garrick Orchard, Christian Mayr, Ryad Benosman, et al.

    Neuromorphic computing is a brain-inspired approach to hardware and algorithm design that efficiently realizes artificial neural networks. Neuromorphic designers apply the principles of biointelligence discovered by neuroscientists to design efficient computational systems, often for applications with size, weight and power constraints. With this research field at a critical juncture, it is crucial to chart the course for the development of future large-scale neuromorphic systems. We describe approaches for creating scalable neuromorphic architectures and identify key features. We discuss potential applications that can benefit from scaling and the main challenges that need to be addressed. Furthermore, we examine a comprehensive ecosystem necessary to sustain growth and the new opportunities that lie ahead when scaling neuromorphic systems. Our work distils ideas from several computing sub-fields, providing guidance to researchers and practitioners of neuromorphic computing who aim to push the frontier forward.


  • Dynamic Neural Fields Accelerator Design for a Millimeter-Scale Tracking System

    by Yuyang Li, Vijay Shankaran Vivekanand, Rajkumar Kubendran, Inhee Lee

    This brief introduces a compact-size hardware accelerator for dynamic neural fields (DNF) used in object tracking. To address the substantial computational workload and memory occupancy associated with conventional DNFs, three key approaches are implemented: kernel size reduction and abstraction, the replacement of sigmoidal functions with comparison operations, and the approximation of rectangular-shaped objects. The design is realized in a 28-nm CMOS process, resulting in a layout with an area of 0.53 mm $^2$ . Simulation results demonstrate that the accelerator processes $256 \times 256$ dynamic vision sensor (DVS) frames at 211 frames per second (fps), with a power consumption of 1.68 mW under such conditions.

  • BioNN: Bio-Mimetic Neural Networks on Hardware Using Nonlinear Multi-Timescale Mixed-Feedback Control for Neuromodulatory Bursting Rhythms

    by Kangni Liu, Shahin Hashemkhani, Jonathan Rubin, Rajkumar Kubendran, Vijay Shankaran Vivekanand, Feng Xiong

    Biological neurons exhibit rich and complex nonlinear dynamics, which are computationally expensive and area/power hungry for hardware implementation. This paper presents a mathematical analysis and hardware realization of neural networks using a nonlinear neuron model that utilizes two excitable systems operating at different timescales. The neuron consists of a mixed-feedback system operating at multiple timescales to exhibit a variety of modalities that resemble the biophysical mechanisms found in neurophysiology. The single neuron dynamics emerge from four voltage-controlled current sources and feature spiking and bursting output modes that can be controlled using tunable parameters. The bifurcation structures of the neuron, modeled as a 4D dynamical system, illustrate the roles of sources acting on different timescales in shaping neural dynamics. A comprehensive understanding of the system’s dynamic behavior is obtained by studying the state space variables and performing bifurcation analysis on the different parameters. The model is implemented to a 1mm x 2mm prototype chip utilizing the 180nm CMOS process. Each neural network consists of 1 isolated test neuron and 5 fully connected neurons using 20 synapses. By carefully selecting bias voltages according to the I-V characterization curves, the neurons are shown to exhibit spike, burst, and burst excitable behavior. Multiple small-scale neural networks with inhibitory or excitatory synapses were verified to achieve coupled rhythms with neuron bursts in-phase or out-of-phase. To demonstrate an application, the generated burst waveforms from the 4-neuron network were used to form a Central Pattern Generator (CPG) for locomotion control of the four legs of the Petoi, a quadruped robot, enabling the bot to jump successfully.


  • Frame of Events: A Low-latency Resource-efficient Approach for Stereo Depth Maps

    by Shanmuga Venkatachalam, Vijay Shankaran Vivekanand, Rajkumar Kubendran

    Computer vision traditionally uses cameras that capture visual information as frames at periodic intervals. On the other hand, Dynamic Vision Sensors (DVS) capture temporal contrast (TC) in each pixel asynchronously and stream them serially. This paper proposes a hybrid approach to generate input visual data as ‘frame of events’ for a stereo vision pipeline. We demonstrate that using hybrid vision sensors that produce frames made up of TC events can achieve superior results in terms of low latency, less compute and low memory footprint as compared to the traditional cameras and the event-based DVS. The frame-of-events approach eliminates the latency and memory resources involved in the accumulation of asynchronous events into synchronous frames, while generating acceptable disparity maps for depth estimation. Benchmarking results show that the frame-of-events pipeline outperforms others with the least average latency per frame of 3.8 ms and least average memory usage per frame of 112.4 Kb, which amounts to 7.32% and 9.75% reduction when compared to traditional frame-based pipeline. Hence, the proposed method is suitable for missioncritical robotics applications that involve path planning and localization mapping in a resource-constrained environment, such as drone navigation and autonomous vehicles.


  • Top Cited Papers
    Open Access

    A compute-in-memory chip based on resistive random-access memory

    by Weier Wan, Rajkumar Kubendran, Clemens Schaefer, Sukru Burc Eryilmaz, Wenqiang Zhang, Dabin Wu, Stephen Deiss, Priyanka Raina, He Qian, Bin Gao, et al.

    Realizing increasingly complex artificial intelligence (AI) functionalities directly on edge devices calls for unprecedented energy efficiency of edge hardware. Compute-in-memory (CIM) based on resistive random-access memory (RRAM)1promises to meet such demand by storing AI model weights in dense, analogue and non-volatile RRAM devices, and by performing AI computation directly within RRAM, thus eliminating power-hungry data movement between separate compute and memory2–5. Although recent studies have demonstrated in-memory matrix-vector multiplication on fully integrated RRAM-CIM hardware6–17, it remains a goal for a RRAM-CIM chip to simultaneously deliver high energy efficiency, versatility to support diverse models and software-comparable accuracy. Although efficiency, versatility and accuracy are all indispensable for broad adoption of the technology, the inter-related trade-offs among them cannot be addressed by isolated improvements on any single abstraction level of the design. Here, by co-optimizing across all hierarchies of the design from algorithms and architecture to circuits and devices, we present NeuRRAM—a RRAM-based CIM chip that simultaneously delivers versatility in reconfiguring CIM cores for diverse model architectures, energy efficiency that is two-times better than previous state-of-the-art RRAM-CIM chips across various computational bit-precisions, and inference accuracy comparable to software models quantized to four-bit weights across various AI tasks, including accuracy of 99.0 percent on MNIST18and 85.7 percent on CIFAR-1019image classification, 84.7-percent accuracy on Google speech command recognition20, and a 70-percent reduction in image-reconstruction error on a Bayesian image-recovery task.


  • High-Speed, Real-Time, Spike-Based Object Tracking and Path Prediction on Google Edge TPU

    by Jonah Sengupta, Rajkumar Kubendran, Emre Neftci, Andreas Andreou

    As computation and advanced high-dimensional signal processing is pushed to edge computational devices, energy efficient, unconventional architectures are needed to ameliorate this growing need. The Google Edge TPU, first used on a Cloud platform, is one such accelerator that is now commercially available for consumer use. Similarly, low-power, data-efficient vision sensors, such as the Dynamic Vision Sensor (DVS), have been developed and commercialized as well to improve upon the large data redundancy seen in these ML applications. This live demonstration is linking these two technologies to benchmark the Coral Edge TPU Board in a high-speed object tracking and prediction application. In comparison to a floating point architecture of similar form factor, the Intel Compute Stick, the Edge TPU has been show to outperform in terms of latency and computational efficiency.


  • Top Cited Papers

    33.1 A 74 TMACS/W CMOS-RRAM Neurosynaptic Core with Dynamically Reconfigurable Dataflow and In-situ Transposable Weights for Probabilistic Graphical Models

    by Weier Wan, Rajkumar Kubendran, S. Burc Eryilmaz, Wenqiang Zhang, Yan Liao, Dabin Wu, Stephen Deiss, Bin Gao, Priyanka Raina, Siddharth Joshi, et al.

    Many powerful neural network (NN) models such as probabilistic graphical models (PGMs) and recurrent neural networks (RNNs) require flexibility in dataflow and weight access patterns as shown in Fig. 33.1.1 Typically, Compute-In-Memory (CIM) designs do not implement such dataflows or do so by replicating circuits at the memory periphery such as ADCs/neurons along both the rows and columns of the memory array, leading to an overhead in operation. This paper describes a CIM architecture implemented in a 130nm CMOS/RRAM process, that delivers the highest reported computational energy-efficiency of 74 tera-multiply-accumulates per second per watt (TMACS/W) for RRAM-based CIM architectures while simultaneously offering dataflow reconfigurability to address the limitations of previous designs. This is made possible through two key features: 1) a runtime reconfigurable dataflow with in-situ access to RRAM array and its transpose for efficient access to NN weights and 2) a voltage sensing stochastic integrate-and-fire analog neuron (I&F) that is reused for correlated double sampling (CDS), stochastic voltage integration, and threshold comparison.