High-speed electro-optic engine for a quantized photonic neural network

Integrated photonic computing possesses remarkable advantages in computational power, becoming a promising candidate for high-performance computing in the post-Moore era. However, integrated photonic computing systems face trade-offs between scale, speed, and precision, hindering their development into high-performance processors. In this work, we propose and experimentally demonstrate a photonic quantized neural network using a large-scale lithium niobate on insulator high-speed electro-optic engine. With the electro-optic engine providing 1.6 TeraOPS computational power, our photonic quantized neural network experiments achieve a high accuracy of 97% in CNN-based MNIST handwritten digit recognition and 80% in GCN-based Cora data node classification, while also realizing over 95% model compression. Based on the quantized input-weight parallel encoding scheme, our high-speed electro-optic engine achieves a 100-fold improvement in system frequency over digital computing architectures and a 10-fold improvement in both modulation speed and compute density over optoelectronic computing architectures, while also achieving competitive energy efficiency. Our experimental findings indicate that our photonic quantized neural network technology paves the way for large-scale and high-speed integrated photonic computing, further promoting its practical application.