Photonic Integrated Circuit for Machine Acceleration: Photonic Tensor Core Processor
Open AccessThe field of machine learning and artificial intelligence (AI) has experienced significant growth and has become increasingly prevalent in modern society. However, the progress of adopting intelligent automation systems is hindered by hardware limitations, including constraints related to throughput, power consumption, and latency. While electronic systems have reached the end of their scaling laws, there is a need for alternative accelerators.Optical co-processors present a potential solution to these hardware limitations by offering a high degree of algorithmic homomorphism. This is achieved by implementing general matrix-matrix multiplication operations through on-the-fly multiplication by electro-optic components, and accumulation operations by photodetectors. Despite this potential, recent emerging photonic AI engines have been found to be cumbersome to program, follow scaling laws that incur overhead, or rely on discretely-packaged photonic components, resulting in decreased performance. In this dissertation, we present two methods for achieving a multiply-accumulate engine on photonic circuits. These methods focus on addressing the different bottlenecks of current photonic AI engines. Specifically, we propose hybrid electronic-photonic tensor processors that offer a versatile, low-overhead, and compact solution for extreme edge AI applications, as well as for machine learning training tasks in the cloud. These novel approaches have the potential to significantly improve the performance and efficiency of AI systems, and thus have broad implications for the field of machine learning and AI.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.