Aishwarya Natarajan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Aishwarya Natarajan has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
- Jim Ignowski
Hewlett Packard Labs, Hewlett Packard Enterprise, Milpitas, USA
4 shared publications - Luca Buonanno
Hewlett Packard Labs, Fort Collins, CO, USA
4 shared publications - Archit Gajjar
Artificial Intelligence Research Lab (AIRL), Hewlett Packard Labs, USA
3 shared publications - Giacomo Pedretti
Hewlett Packard Labs, Hewlett Packard Enterprise, Milpitas, USA
3 shared publications - Jennifer Hasler
Electrical and Computer Engineering (ECE), Georgia Institute of Technology, Atlanta, GA 30332, USA
3 shared publications - Lei Zhao
Artificial Intelligence Research Lab (AIRL), Hewlett Packard Labs, USA
3 shared publications
Latest Publications
- Open Access
- 23 October 2026
- Published by Springer Nature in npj Unconventional Computing
- Vol. 3 (1) , 56
SEPhIA: spiking electro-photonic integrated tiled architecture leveraging multi-wavelength lasers for scalable optical neuromorphic computing
Research into optical spiking neural networks (SNNs) has primarily focused on spiking devices, networks of excitable lasers or numerical modeling of large architectures, often overlooking key constraints such as limited optical power, crosstalk and footprint. We introduce SEPhIA, a photonic-electronic, multi-tiled SNN architecture emphasizing implementation feasibility and realistic scaling. SEPhIA leverages microring modulators (MRMs) and multi-wavelength sources to achieve effective $$ < 1$$ laser-per-spiking neuron ratio. We validate SEPhIA at both device and architecture levels by time‑domain co‑simulating excitable CMOS‑MRM coupled circuits and by devising a physics‑aware, trainable optoelectronic SNN model, with both approaches utilizing experimentally derived device parameters. The multi-layer optoelectronic SNN achieves classification accuracies over 90% on a four-class spike-encoded dataset, closely comparable to software models. A design space study further quantifies how photonic device parameters impact SNN performance under constrained signal-to-noise conditions. SEPhIA offers a scalable, expressive, physically grounded solution for neuromorphic photonic computing, capable of addressing spike-encoded tasks.- 23 October 2026
- Published by Springer Nature in npj Unconventional Computing
- Vol. 3 (1) , 56
Open AccessSEPhIA: spiking electro-photonic integrated tiled architecture leveraging multi-wavelength lasers for scalable optical neuromorphic computing
Research into optical spiking neural networks (SNNs) has primarily focused on spiking devices, networks of excitable lasers or numerical modeling of large architectures, often overlooking key constraints such as limited optical power, crosstalk and footprint. We introduce SEPhIA, a photonic-electronic, multi-tiled SNN architecture emphasizing implementation feasibility and realistic scaling. SEPhIA leverages microring modulators (MRMs) and multi-wavelength sources to achieve effective $$ < 1$$ laser-per-spiking neuron ratio. We validate SEPhIA at both device and architecture levels by time‑domain co‑simulating excitable CMOS‑MRM coupled circuits and by devising a physics‑aware, trainable optoelectronic SNN model, with both approaches utilizing experimentally derived device parameters. The multi-layer optoelectronic SNN achieves classification accuracies over 90% on a four-class spike-encoded dataset, closely comparable to software models. A design space study further quantifies how photonic device parameters impact SNN performance under constrained signal-to-noise conditions. SEPhIA offers a scalable, expressive, physically grounded solution for neuromorphic photonic computing, capable of addressing spike-encoded tasks. - 28 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. PP (02780070) , 1
NL-DPE: An Analog In-memory Non-Linear Dot Product Engine for Efficient CNN and LLM Inference
Resistive Random Access Memory (RRAM) based in-memory computing (IMC) accelerators offer significant performance and energy advantages for deep neural networks (DNNs), but face three major limitations: (1) they support only static dot-product operations and cannot accelerate arbitrary non-linear functions or data-dependent multiplications essential to modern LLMs; (2) they demand large, power-hungry analog-to-digital converter (ADC) circuits; and (3) mapping model weights to device conductance introduces errors from cell nonidealities. These challenges hinder scalable and accurate IMC acceleration as models grow. We propose NL-DPE, a Non-Linear Dot Product Engine that overcomes these barriers. NL-DPE augments crosspoint arrays with RRAM-based Analog Content Addressable Memory (ACAM) to execute arbitrary non-linear functions and data-dependent matrix multiplications in the analog domain by transforming them into decision trees, fully eliminating ADCs. To address device noise, NL-DPE uses software-based Noise Aware Fine-tuning (NAF), requiring no in-device calibration. Experiments show that NL-DPE delivers 28× energy efficiency and 249× speedup over a GPU baseline, and 22× energy efficiency and 245× speedup over existing IMC accelerators, while maintaining high accuracy.
- 28 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. PP (02780070) , 1
NL-DPE: An Analog In-memory Non-Linear Dot Product Engine for Efficient CNN and LLM Inference
Resistive Random Access Memory (RRAM) based in-memory computing (IMC) accelerators offer significant performance and energy advantages for deep neural networks (DNNs), but face three major limitations: (1) they support only static dot-product operations and cannot accelerate arbitrary non-linear functions or data-dependent multiplications essential to modern LLMs; (2) they demand large, power-hungry analog-to-digital converter (ADC) circuits; and (3) mapping model weights to device conductance introduces errors from cell nonidealities. These challenges hinder scalable and accurate IMC acceleration as models grow. We propose NL-DPE, a Non-Linear Dot Product Engine that overcomes these barriers. NL-DPE augments crosspoint arrays with RRAM-based Analog Content Addressable Memory (ACAM) to execute arbitrary non-linear functions and data-dependent matrix multiplications in the analog domain by transforming them into decision trees, fully eliminating ADCs. To address device noise, NL-DPE uses software-based Noise Aware Fine-tuning (NAF), requiring no in-device calibration. Experiments show that NL-DPE delivers 28× energy efficiency and 249× speedup over a GPU baseline, and 22× energy efficiency and 245× speedup over existing IMC accelerators, while maintaining high accuracy.
- Open Access
- 27 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (1) , 1-26
Azure-Lily: An FPGA Architecture with Analog IMC Engines for Efficient AI
Modern AI models place heavy demands on compute resources, underscoring the importance of hardware accelerators that can balance performance, energy, and flexibility. The ever-growing demand for AI computing, coupled with slowing performance gains in chip manufacturing, has heightened the role of FPGA-based accelerators due to their rapid adaptability, reprogrammability, and support for custom parallel data flows. In this work, we introduce a domain-optimized FPGA architecture designed for deep neural network (DNN) inference by embedding analog in-memory compute blocks, specifically, RRAM-based Dot Product Engines (DPE), directly into the fabric. These engines perform multiply-accumulate operations within an RRAM-based crossbar, minimizing data movement through the FPGA routing fabric, while enhancing the compute throughput of the FPGA significantly, overall improving energy efficiency of mapping DNNs on FPGAs. We evaluate our architecture using the Verilog-to-Routing (VTR) framework. We simulate a novel 22 nm architecture, and employ a custom event-driven simulator to evaluate its performance. Our design achieves an average 6.58× reduction in latency, 5.42× throughput improvement, and 8,741× energy efficiency improvement when compared to state-of-the-art FPGA implementations across multiple DNNs (LeNet, ResNet, VGG).
- 27 March 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (1) , 1-26
Open AccessAzure-Lily: An FPGA Architecture with Analog IMC Engines for Efficient AI
Modern AI models place heavy demands on compute resources, underscoring the importance of hardware accelerators that can balance performance, energy, and flexibility. The ever-growing demand for AI computing, coupled with slowing performance gains in chip manufacturing, has heightened the role of FPGA-based accelerators due to their rapid adaptability, reprogrammability, and support for custom parallel data flows. In this work, we introduce a domain-optimized FPGA architecture designed for deep neural network (DNN) inference by embedding analog in-memory compute blocks, specifically, RRAM-based Dot Product Engines (DPE), directly into the fabric. These engines perform multiply-accumulate operations within an RRAM-based crossbar, minimizing data movement through the FPGA routing fabric, while enhancing the compute throughput of the FPGA significantly, overall improving energy efficiency of mapping DNNs on FPGAs. We evaluate our architecture using the Verilog-to-Routing (VTR) framework. We simulate a novel 22 nm architecture, and employ a custom event-driven simulator to evaluate its performance. Our design achieves an average 6.58× reduction in latency, 5.42× throughput improvement, and 8,741× energy efficiency improvement when compared to state-of-the-art FPGA implementations across multiple DNNs (LeNet, ResNet, VGG).
- Open Access
Memristive tabular variational autoencoder for compression of analog data in high energy physics
We present an implementation of edge AI to compress data on an in-memory analog content-addressable memory (ACAM) device. A variational autoencoder is trained on a simulated sample of energy measurements from incident high-energy electrons on a generic three-layer scintillator-based calorimeter. The encoding part is distilled into tabular format by regressing the latent space variables using decision trees, which is then programmed on a memristor-based ACAM. In real-time, the ACAM compresses 48 continuously valued incoming energies measured by the calorimeter sensors into the latent space, achieving a compression factor of 12x, which is transmitted off-detector for decompression. The performance result of the ACAM, obtained using the Structural Simulation Toolkit, the SST open source framework, gives a latency value of 24 ns and a throughput of 330M compressions per second, i.e., 3 ns between successive inputs, and an average energy consumption of 4.1 nJ per compression.
Open AccessMemristive tabular variational autoencoder for compression of analog data in high energy physics
We present an implementation of edge AI to compress data on an in-memory analog content-addressable memory (ACAM) device. A variational autoencoder is trained on a simulated sample of energy measurements from incident high-energy electrons on a generic three-layer scintillator-based calorimeter. The encoding part is distilled into tabular format by regressing the latent space variables using decision trees, which is then programmed on a memristor-based ACAM. In real-time, the ACAM compresses 48 continuously valued incoming energies measured by the calorimeter sensors into the latent space, achieving a compression factor of 12x, which is transmitted off-detector for decompression. The performance result of the ACAM, obtained using the Structural Simulation Toolkit, the SST open source framework, gives a latency value of 24 ns and a throughput of 330M compressions per second, i.e., 3 ns between successive inputs, and an average energy consumption of 4.1 nJ per compression.
- Open Access
- 27 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 179
Enhancing FPGAs with Analog In-Memory Computing Macros
While the AI computing needs are ever-increasing and the innovation in models generates tens of new architectures yearly, the performance gain from improvements in chip manufacturing has slowed down. Within this context, FPGA-based accelerators play a fundamental role. FPGAs are the backbone of specialized architectures, their reconfigurability being the key differentiation that enables an effective design space exploration. At the same time, to overcome the limitations induced by the memory bottleneck, the computing architectures community has proposed the in-memory computing paradigm: storage and computations are both performed in non-volatile memory devices. We propose the enhancement of FPGAs with in-memory computing macros, specifically analog Dot Product Engines, based on non-volatile ReRAM devices. We designed a novel 40\:nm,\break 23.7\:mm×23.7\:mm architecture using the VTR framework and used a custom event-driven simulator to estimate its projected performances. When comparing its performances to the state-of-the-art, the in-memory computing enhanced FPGA achieves 154\:TOPS, a ×234 throughput improvement, and 2.08\:TOPS/W, a ×143 energy efficiency improvement with respect to published results for FPGA implementations of AI models. This paper intends to demonstrate to the FPGA community the breakthrough advantages in energy efficiency and throughput derived from the integration of Dot Product Engines in the FPGA fabric, aiming at fulfilling the exponentially increasing computing requirements of the AI hardware and HPC industry.
- 27 February 2025
- Published by Association for Computing Machinery (ACM)
- p. 179
Open AccessEnhancing FPGAs with Analog In-Memory Computing Macros
While the AI computing needs are ever-increasing and the innovation in models generates tens of new architectures yearly, the performance gain from improvements in chip manufacturing has slowed down. Within this context, FPGA-based accelerators play a fundamental role. FPGAs are the backbone of specialized architectures, their reconfigurability being the key differentiation that enables an effective design space exploration. At the same time, to overcome the limitations induced by the memory bottleneck, the computing architectures community has proposed the in-memory computing paradigm: storage and computations are both performed in non-volatile memory devices. We propose the enhancement of FPGAs with in-memory computing macros, specifically analog Dot Product Engines, based on non-volatile ReRAM devices. We designed a novel 40\:nm,\break 23.7\:mm×23.7\:mm architecture using the VTR framework and used a custom event-driven simulator to estimate its projected performances. When comparing its performances to the state-of-the-art, the in-memory computing enhanced FPGA achieves 154\:TOPS, a ×234 throughput improvement, and 2.08\:TOPS/W, a ×143 energy efficiency improvement with respect to published results for FPGA implementations of AI models. This paper intends to demonstrate to the FPGA community the breakthrough advantages in energy efficiency and throughput derived from the integration of Dot Product Engines in the FPGA fabric, aiming at fulfilling the exponentially increasing computing requirements of the AI hardware and HPC industry.
- Open Access
- 24 November 2018
- Published by MDPI AG in Journal of Low Power Electronics and Applications
- Vol. 8 (4) , 47
Enabling Energy-Efficient Physical Computing through Analog Abstraction and IP Reuse
This paper shows the first step in analog (and mixed signal) abstraction utilized in large-scale Field Programmable Analog Arrays (FPAA), encoded in the open-source SciLab/Xcos based toolset. Having any opportunity of a wide-scale utilization of ultra-low power technology both requires programmability/reconfigurability as well as abstractable tools. Abstraction is essential both make systems rapidly, as well as reduce the barrier for a number of users to use ultra-low power physical computing techniques. Analog devices, circuits, and systems are abstractable and retain their energy efficient opportunities compared with custom digital hardware. We will present the analog (and mixed signal) abstraction developed for the open-source toolkit used for the SoC FPAAs. Abstraction of Blocks in the FPAA block library makes the SoC FPAA ecosystem accessible to system-level designers while still enabling circuit designers the freedom to build at a low level. Multiple working test cases of various levels of complexity illustrate the analog abstraction capability. The FPAA block library provides a starting point for discussing the fundamental block concepts of analog computational approaches.
- 24 November 2018
- Published by MDPI AG in Journal of Low Power Electronics and Applications
- Vol. 8 (4) , 47
Open AccessEnabling Energy-Efficient Physical Computing through Analog Abstraction and IP Reuse
This paper shows the first step in analog (and mixed signal) abstraction utilized in large-scale Field Programmable Analog Arrays (FPAA), encoded in the open-source SciLab/Xcos based toolset. Having any opportunity of a wide-scale utilization of ultra-low power technology both requires programmability/reconfigurability as well as abstractable tools. Abstraction is essential both make systems rapidly, as well as reduce the barrier for a number of users to use ultra-low power physical computing techniques. Analog devices, circuits, and systems are abstractable and retain their energy efficient opportunities compared with custom digital hardware. We will present the analog (and mixed signal) abstraction developed for the open-source toolkit used for the SoC FPAAs. Abstraction of Blocks in the FPAA block library makes the SoC FPAA ecosystem accessible to system-level designers while still enabling circuit designers the freedom to build at a low level. Multiple working test cases of various levels of complexity illustrate the analog abstraction capability. The FPAA block library provides a starting point for discussing the fundamental block concepts of analog computational approaches.
- 13 July 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Circuits and Systems
- Vol. 12 (4) , 918-926
Hodgkin–Huxley Neuron and FPAA Dynamics
We present the experimental silicon results on the dynamics of a Hodgkin-Huxley neuron implemented on a reconfigurable platform. The circuit has been inspired by the similarity between biology and silicon, by modeling ion channels and their time constants. Another significant motivation behind this paper is to make the system available to circuit designers as well as users in the neuroscience community. The open-source tool infrastructure and a remote system ease the accessibility of our system to a number of users. We demonstrate the reproducibility of the results by replicating the dynamics across different boards along with responses from different inputs and with different parameters. The reconfigurability enables one to make use of a single primary design to obtain a variety of results. The measurements are taken from the system compiled on a field programmable analog array fabricated on a 350-nm process.
- 13 July 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Circuits and Systems
- Vol. 12 (4) , 918-926
Hodgkin–Huxley Neuron and FPAA Dynamics
We present the experimental silicon results on the dynamics of a Hodgkin-Huxley neuron implemented on a reconfigurable platform. The circuit has been inspired by the similarity between biology and silicon, by modeling ion channels and their time constants. Another significant motivation behind this paper is to make the system available to circuit designers as well as users in the neuroscience community. The open-source tool infrastructure and a remote system ease the accessibility of our system to a number of users. We demonstrate the reproducibility of the results by replicating the dynamics across different boards along with responses from different inputs and with different parameters. The reconfigurability enables one to make use of a single primary design to obtain a variety of results. The measurements are taken from the system compiled on a field programmable analog array fabricated on a 350-nm process.
- Open Access
- 30 August 2017
- Published by MDPI AG in Journal of Low Power Electronics and Applications
- Vol. 7 (3) , 21
Models and Techniques for Temperature Robust Systems on a Reconfigurable Platform
This paper investigates the variability of various circuits and systems over temperature and presents several methods to improve their performance over temperature. The work demonstrates use of large scale reconfigurable System-On-Chip (SOC) for reducing the variability of circuits and systems compiled on a Floating Gate (FG) based Field Programmable Analog Array (FPAA). Temperature dependencies of circuits are modeled using an open-source simulator built in the Scilab/XCOS environment and the results are compared with measurement data obtained from the FPAA. This comparison gives further insight into the temperature dependence of various circuits and signal processing systems and allows us to compensate as well as predict their behavior. Also, the work presents several different current and voltage references that could help in reducing the variability caused due to changes in temperature. These references are standard blocks in the Scilab/Xcos environment that could be easily compiled on the FPAA. An FG based current reference is then used for biasing a Vector Matrix Multiplication (VMM) circuit and a second order bandpass filter to demonstrate the compilation and usage of these voltage/current reference in a reconfigurable fabric. The large scale FG FPAA presented here is fabricated in a 350 nm CMOS process.- 30 August 2017
- Published by MDPI AG in Journal of Low Power Electronics and Applications
- Vol. 7 (3) , 21
Open AccessModels and Techniques for Temperature Robust Systems on a Reconfigurable Platform
This paper investigates the variability of various circuits and systems over temperature and presents several methods to improve their performance over temperature. The work demonstrates use of large scale reconfigurable System-On-Chip (SOC) for reducing the variability of circuits and systems compiled on a Floating Gate (FG) based Field Programmable Analog Array (FPAA). Temperature dependencies of circuits are modeled using an open-source simulator built in the Scilab/XCOS environment and the results are compared with measurement data obtained from the FPAA. This comparison gives further insight into the temperature dependence of various circuits and signal processing systems and allows us to compensate as well as predict their behavior. Also, the work presents several different current and voltage references that could help in reducing the variability caused due to changes in temperature. These references are standard blocks in the Scilab/Xcos environment that could be easily compiled on the FPAA. An FG based current reference is then used for biasing a Vector Matrix Multiplication (VMM) circuit and a second order bandpass filter to demonstrate the compilation and usage of these voltage/current reference in a reconfigurable fabric. The large scale FG FPAA presented here is fabricated in a 350 nm CMOS process. - 6 January 2017
- Published by Springer Nature in Analog Integrated Circuits and Signal Processing
- Vol. 91 (1) , 119-130
Modeling, simulation and implementation of circuit elements in an open-source tool set on the FPAA
- 6 January 2017
- Published by Springer Nature in Analog Integrated Circuits and Signal Processing
- Vol. 91 (1) , 119-130
Modeling, simulation and implementation of circuit elements in an open-source tool set on the FPAA