MDTransformer: A Hardware-Software Co-Design of Mode-Division Photonic Transformer Accelerator with Inverse-Designed Coherent Crossbar
MDTransformer uses mode-division photonics to run four optical compute lanes per waveguide while avoiding multi-wavelength light sources.
The paper proposes a photonic transformer accelerator built around inverse-designed multi-mode components and a compact mode-division photonic tensor core. It uses coherent detection and IQ modulation to support complex-valued arithmetic for transformer operations. The authors report sub-4-bit effective analog precision, inter-modal crosstalk below -30 dB, and single-laser operation at 1550 nm. Across DeiT and BERT workloads, they claim 40.4% area reduction, 63.6% power saving, and 40.6% energy saving versus a prior photonic transformer accelerator, with comparable latency. Source: ArXiv · AI/CL/LG's note.
The paper proposes a photonic transformer accelerator built around inverse-designed multi-mode components and a compact mode-division photonic tensor core. It uses coherent detection and IQ modulation to support complex-valued arithmetic for transformer operations. The authors report sub-4-bit effective analog precision, inter-modal crosstalk below -30 dB, and single-laser operation at 1550 nm. Across DeiT and BERT workloads, they claim 40.4% area reduction, 63.6% power saving, and 40.6% energy saving versus a prior photonic transformer accelerator, with comparable latency. Source: ArXiv · AI/CL/LG's note.
score 4