Grants and Contracts Details
Abstract [subject to minor changes] Overview: In the past decade, deep neural networks (DNNs) have demonstrated record-breaking performance on various artificial intelligence tasks. This has motivated the development of novel neural network accelerators (NNAs) to improve the inference and training performance of DNNs. However, the escalating model scales and computation demands of DNNs present substantial technical challenges to traditional electrical NNA platforms. As a compelling alternative, optical NNAs (ONNAs) have attracted increasing attention. Detailed analysis of early ONNA prototypes has shown that ONNAs can potentially provide disruptive throughput in the hundreds of tera operations (TOPs)/s with energy efficiencies of sub-fJ/OP, which cannot be achieved by traditional electrical NNAs due to their poor scalability. Despite these efficiency benefits, ONNAs face several daunting challenges that hinder their immediate widespread adoption. First, ONNAs suffer from very low area-efficiency, which restricts their applicability in the cost-tolerant high-end systems only. Second, the presence of several crippling factors, such as unavoidable optical signal losses, fabrication-process and on-chip temperature variations, and stringent optical power budget, severely confines the scalability of the weighting precision and fan in of ONNAs. This makes ONNAs deficiently less adaptive to advanced model scaling. Third, the state-of-the-art implementations of ONNAs cannot support variable weighting precision, which makes them incompatible with modern mixed-precision based high-accuracy training algorithms. The objective of proposed research project is to overcome these challenges by aiming to design novel and potentially groundbreaking ONNA architectures that can synergistically combine the energy-efficiency and latency benefits of optical computing with the area-efficiency, precision reconfigurability, and error resiliency of stochastic computing. To achieve this goal, this project will undertake the following exploratory, yet highly transformative, research tasks in a highly interactive manner across different layers of the hardware design stack: (1) Design and fabricate low-overhead optical logic gates and circuits to realize stochastic arithmetic based ONNA architectures, which can implement various compute-heavy neural network functions, such as multiply-accumulate, pooling, and non-linear activation, by transforming them into simple bit-wise, variable-precision logical functions; (2) Devise cross-layer strategies for efficient system-level integration and dynamic manipulation of stochastic correlation to simultaneously maximize the throughput, accuracy, and error resilience of the designed ONAA architectures; (3) Characterize and validate the designed ONAA architectures using fabrication based prototyping at the circuit-level and cycle- accurate simulations at the system-level. These research tasks will employ the tools and facilities available at the University of Kentucky, through the Kentucky Multiscale NNCI node. Intellectual Merit: The proposed research is important and radically transformative because it aims to merge the disciplines of stochastic computing (SC) and optical computing (OC) in a highly synergistic manner. Such synergistic integration of SC and OC, which has never been explored before, will lay down a blueprint for realizing a new class of ONNA architectures that will surpass existing electrical and optical NNA architectures in terms of speed, energy-efficiency, area-efficiency, accuracy, error resilience, and reconfigurability. Such confluence of SC and OC will also revolutionize the frontiers of SC and OC, both individually as well holistically. On one hand, it will enable dramatic reduction in the inherently high latency of SC by enabling the use of the extreme-scale wavelength-level parallelism of OC for implementing various traditionally bit-serial stochastic functions in a bit-parallel manner. On the other hand, it will help to significantly extenuate the crosstalk noise and reliability issues that have traditionally crippled OC by exploiting the inherent error resilience of SC. Broader Impacts: The highly efficient ONNA architectures that emerge from the proposed research will be at the heart of the revolutionary digital intelligence of the next generation, which will be the driving force for technological advances across various application domains, including the medical, aerospace, consumer, and automotive domains. Moreover, by catalyzing new outreach opportunities for exposing undergraduate and graduate students to the diverse aspects of stochastic arithmetic, correlation manipulation, probability theory, optical computing, nanofabrication, deep neural networks, and electro- optical characterization and testing, the proposed research will contribute towards an agile, high-tech workforce that will maintain continued US leadership in technological innovation.
|Effective start/end date||10/1/21 → 9/30/23|
- National Science Foundation: $299,735.00
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.