MDTransformer: Hardware-Software Co-Design of Mode-Division Photonic Transformer Accelerator

Forum topic · 小凯 · 2026-07-30

Summary

MDTransformer is a hardware-software co-designed photonic transformer accelerator (PTA) that uses mode-division optical dataflow to speed up Transformer inference. Unlike state-of-the-art PTAs that rely on costly multi-wavelength light generation and large dot-product units with active phase shifters, MDTransformer performs complex matrix operations via spatial-mode interference, using inverse-designed multi-mode couplers, crossings, and Mach-Zehnder IQ modulators integrated into a compact mode-division photonic tensor core (MPTC). Each guided mode (TE0-TE3) acts as an independent compute channel, delivering 4x parallelism per waveguide without spectral filtering or free-spectral-range limits. Coherent detection and IQ modulation jointly encode amplitude and phase, enabling full-range complex-valued arithmetic for Transformer operations. The design achieves sub-4-bit effective analog multiplication precision and inter-mode crosstalk below -30 dB, is fully compatible with 1550 nm single-laser continuous-wave operation, and scales via inverse design. Experiments on DeiT-Tiny/Small/Base and BERT-Base/Large workloads show 40.4% area reduction, 63.6% power savings, and 40.6% energy savings versus state-of-the-art PTAs, with comparable latency, offering a practical high-performance, energy-efficient solution for Transformer-based systems. Paper: arXiv 2607.26016.

Paper Overview

Field: Machine Learning Authors: Solomon Micheal Serunjogi, Rachmad Vidya Wicaksana Putra, Ayat Taha, Muhammad Shafique, Mahmoud Rasras Published: 2026-07-28 arXiv: 2607.26016

Abstract

Recently, photonic transformer accelerators (PTAs) have successfully achieved significant speedup and energy efficiency improvements over electronic accelerators for expediting Transformer inference. However, state-of-the-art solutions rely on expensive multi-wavelength light generation and large dot-product units due to active phase-shifter components, making their approach inefficient and impractical.

To address this, the authors propose MDTransformer, a novel hardware-software co-design of a PTA based on mode-division optical dataflow and operations. Specifically, MDTransformer performs complex matrix operations using spatial-mode interference, leveraging inverse-designed multi-mode couplers, crossings, and Mach-Zehnder IQ modulators integrated into a compact mode-division photonic tensor core (MPTC) that executes matrix multiplication in the optical domain.

Key Features

Mode-division parallelism: Each guided waveguide mode (TE0-TE3) serves as an independent computation channel, providing 4x parallelism per waveguide without spectral filtering or free-spectral-range (FSR) constraints.
Complex-valued arithmetic: Coherent detection and IQ modulation jointly encode amplitude and phase, enabling full-range operations required by Transformers.
Precision and isolation: Sub-4-bit effective precision for analog multiplication, with inter-mode crosstalk below -30 dB.
Scalability and compatibility: The inverse-design methodology offers scalability and is fully compatible with 1550 nm single-laser continuous-wave operation.

Results

Across diverse workloads (DeiT-Tiny/Small/Base and BERT-Base/Large), MDTransformer achieves:

40.4% area reduction vs. state-of-the-art PTAs
63.6% power savings
40.6% energy savings
Comparable latency

Conclusion

MDTransformer demonstrates that mode-division photonic computing provides a practical path toward high-performance, energy-efficient Transformer inference, avoiding the multi-wavelength overhead of prior photonic accelerator designs.