Research Projects
NEOCNN: NTT-Enabled Optical Convolution Neural Network Accelerator
ICS 2024This project introduces NEOCNN, an optical accelerator for convolutional neural networks (CNNs) that leverages the Number Theoretic Transform (NTT) for efficient computation. The research addresses the growing demand for high-speed, low-latency processing of CNNs, particularly in applications requiring real-time performance. NEOCNN utilizes the inherent parallelism of optics to perform NTT-based convolutions, which can significantly reduce the computational complexity and energy consumption compared to traditional electronic implementations. The expected outcomes of this project include a novel architecture for optical CNN acceleration, a comprehensive performance analysis, and a demonstration of its potential for next-generation AI hardware.
Related Links
FIONA: Photonic-Electronic CoSimulation Framework and Transferable Prototyping for Photonic Accelerator
ICCAD 2023FIONA is a comprehensive co-simulation framework and prototyping platform for photonic accelerators. This project aims to bridge the gap between the photonics and computer architecture communities by providing a unified environment for designing, simulating, and prototyping photonic-electronic systems. FIONA addresses the challenges of co-designing hardware and software for photonic accelerators by offering a flexible and extensible framework that supports various levels of abstraction, from device physics to system-level performance. The key findings of this work include a novel co-simulation methodology, a set of reusable hardware and software components, and a demonstration of the framework's effectiveness through the design and implementation of a prototype photonic accelerator.
Related Links
RONet: Scaling GPU System with Silicon Photonic Chiplet
ICCAD 2023This project explores the use of silicon photonic chiplets to address the scalability challenges of multi-GPU systems. RONet proposes a novel architecture that leverages the high bandwidth and low latency of optical interconnects to create a scalable and efficient communication fabric for GPUs. The research investigates the design of the photonic chiplet, the network-on-chip architecture, and the software stack required to enable seamless integration with existing GPU systems. The goal is to demonstrate a significant improvement in performance and energy efficiency for large-scale deep learning and high-performance computing workloads.
Related Links
PhotonNTT: Energy-Efficient Parallel Photonic Number Theoretic Transform Accelerator
DATE 2024This project presents PhotonNTT, a photonic accelerator for the Number Theoretic Transform (NTT), a key component in post-quantum cryptography. The research focuses on designing an energy-efficient and highly parallel architecture that can accelerate NTT computations for large-scale cryptographic applications. PhotonNTT leverages the unique properties of photonics to perform the complex multiplications and additions required by the NTT algorithm with high speed and low power consumption. The work includes a detailed analysis of the photonic device design, the system architecture, and a comparison with state-of-the-art electronic implementations.
Related Links
Towards Scalable GPU System with Silicon Photonic Chiplet
DATE 2024This project investigates the use of silicon photonic chiplets to create scalable and high-performance GPU systems. The research addresses the limitations of electrical interconnects in large-scale GPU clusters and proposes a novel architecture based on optical communication. The work includes the design of the photonic chiplet, the integration with the GPU, and the development of a communication protocol that can efficiently handle the high-bandwidth traffic between GPUs. The goal is to demonstrate a significant improvement in the performance and scalability of GPU systems for a wide range of applications, from scientific computing to machine learning.