S. Ahmed, R. Baghdadi, Mikhail Bernadskiy, Nate Bowman, R. Braid, Jim Carr, Chen Chen, Pietro Ciccarella et al.
Proposes a universal photonic integrated AI accelerator architecture based on reconfigurable optical circuits that can accelerate various neural networks.
Electronic AI accelerators face slowdown of Moore's law and energy consumption limits, while existing optical computing research is optimized for specific tasks (e.g., matrix multiplication) and lacks versatility. An optical hardware platform that can efficiently support diverse AI models is needed.
The proposed architecture uses reconfigurable optical circuits based on Mach-Zehnder interferometer arrays to perform matrix operations and nonlinear activation functions in the optical domain. By adjusting the phase modulators of the circuit, it can be dynamically reconfigured for different neural network structures (CNN, RNN, Transformer, etc.). Additionally, on-chip lasers and detectors are integrated to implement all-optical signal processing, and an electronic control interface supports training and inference.
Simulation results show that the proposed accelerator achieves over 10x energy efficiency improvement compared to electronic GPUs, and operates without accuracy loss on various benchmark tasks (CIFAR-10, LSTM language model). This study is the first to demonstrate the versatility of optical AI accelerators, providing an important milestone for next-generation low-power AI hardware.