Large-scale photonic chiplet Taichi empowers 160-TOPS/W artificial general intelligence

The everlasting pursuit for a more powerful artificial general intelligence (AGI) continuously demands for higher computing performance. Breaking the restrictions from the Moore’s Law, integrated photonic neural networks have shown great potential in terms of superior processing speed and high energy-efficiency. However, the capacity and scalability are restricted by the unavoidable and time-varying on-chip errors that only simple tasks and shallow models (e.g., 10-category digits classification with thousand-parameter-level networks) can be on-chip realized optically. To support modern AGIs with optical computing, we innovate Taichi, large-scale photonic chiplets owning millions-of-neuron on-chip computing capability with 160 TOPS/W energy efficiency. For the first time, Taichi experimentally realized on-chip optical neural networks (ONNs) with over 13.96M neurons for thousand-category-level classification (tested 91.89% of accuracy for 1623-category Omniglot dataset) and high-fidelity artificial intelligence-generated content (AIGC) with state-of-the-art performances, while achieving up to 2 orders of magnitude improvement in computing efficiency than state-of-the-art AI chips. Based on an integrated diffractive-interference-hybrid design and a general distributed computing architecture, Taichi paves its own way for large-scale on-chip ONN models and advanced tasks, further exploiting the flexibility and potential of optical devices, which will lead to a giant leap in scalability, accuracy and efficiency to support modern AGI.

Methods
The preprocessed dataset that modified from the original ones. We resized/cropped the dimension of the images and re-origanized the training-testing splitting in the original datasets, to fit the data requirements by our optical computing hardware Taichi chiplets.