arXiv Paper: On-chip 1 TOPS Hyperdimensional Photonic Tensor Core Achieved with WDM Silicon Photonic Crossbar

arXiv (学術論文プレプリント) そ 他

Overview

This paper demonstrates an on-chip 0.96 TOPS (tera-operations per second) hyperdimensional photonic tensor core utilizing a Time-, Spatial-, and Wavelength-Division Multiplexing (TSWDM) silicon photonic crossbar. This novel architecture unfolds large matrix-vector and tensor-vector products in time, distributing computational loads across different spatial and wavelength channels. Experimental operation of a 4-channel 2-input TSWDM crossbar incorporating 56 GHz electroabsorption modulators (EAMs) and a 4-channel integrated multiplexing stage was shown, with its AI accelerator performance evaluated on the Iris dataset classification.

In Depth

The Potential of Optical Computing in AI Acceleration

AI workloads, such as deep learning, necessitate the execution of an immense number of matrix operations at high speed and low power. Traditional electronic AI accelerators face physical limits in power consumption and computational speed for this task, making optical computing, which performs calculations using light, a promising alternative. Particularly, silicon photonics integrated with Wavelength Division Multiplexing (WDM) technology holds the potential to significantly enhance computational power and power efficiency.

Realization of On-chip 1 TOPS Photonic Tensor Core

This paper demonstrates a “hyperdimensional photonic tensor core” with a high computational power of 0.96 TOPS (tera-operations per second) on-chip, utilizing a Time-, Spatial-, and Wavelength-Division Multiplexing (TSWDM) silicon photonic crossbar. This is a groundbreaking achievement in performing AI acceleration in the optical domain at the chip level. This new architecture features the following key characteristics:

• High Efficiency Through Multiplexing: By distributing the computational load of large matrix-vector and tensor-vector products across time, different spatial channels, and wavelength channels, it enables efficient parallel processing.
• High-Performance Components: Experimental operation of a 4-channel 2-input TSWDM crossbar incorporating 56 GHz electroabsorption modulators (EAMs) and a 4-channel integrated multiplexing stage was shown. EAMs are critical devices for achieving high-speed optical modulation.
• AI Performance Evaluation: Its performance as an AI accelerator was evaluated on the Iris dataset classification task, achieving a high experimental accuracy of 93.3% at data rates of 4×10 to 4×30 GBd (gigabaud per second). When increasing the data rate to 4×60 GBd, the accuracy was 83.3%.

Technical Significance and Industry Impact

This achievement holds the potential to significantly enhance the performance of AI acceleration through optical computing at the chip level. The introduction of WDM reduces laser operating power and increases the possibility of building photonic accelerators with POPS (Peta-operations per second) level computational throughput. For the power consumption challenges associated with increasing AI model computational load, power-efficient and high-speed acceleration via optical computing contributes to reducing data center operational costs and improving performance. While currently a laboratory-level achievement, requiring further application to larger AI workloads, reliability, and establishment of mass production technologies, this research indicates an important direction for shaping the future of AI computing.