Reproduce “Lightening-Transformer: A Dynamically-operated Optically-interconnected Photonic Transformer Accelerator” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hanqing Zhu, Jiaqi Gu, Hanrui Wang, Zixuan Jiang, Zhekai Zhang, Rongxing Tang, Chenghao Feng, Song Han and 2 more
Lightening-Transformer introduces the first photonic accelerator capable of efficiently processing attention-based Transformers, including large language models, by handling dynamically-encoded, full-range optical operands. The system demonstrates a 2-3 orders of magnitude reduction in Energy-Delay Product (EDP) compared to state-of-the-art electronic accelerators, alongside superior energy efficiency and latency while preserving accuracy.
Key insights
- A holistic co-design approach, spanning optical device physics to system architecture, is crucial for addressing the fundamental limitations and system-level bottlenecks of photonic AI accelerators.
- Coherent optical interference combined with balanced photodetection provides an effective mechanism for handling full-range, dynamic operands in photonic computing, significantly expanding its applicability.
- Strategic use of optical interconnects for operand broadcast and analog-domain temporal accumulation dramatically reduces energy and latency overheads associated with electrical-to-optical conversions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.