ARTFEED — Contemporary Art Intelligence

MDTransformer: Photonic Accelerator for Transformer Inference

ai-technology · 2026-07-29

MDTransformer represents a co-design of hardware and software for a photonic transformer accelerator (PTA) that utilizes mode-division optical dataflow and operations. It executes intricate matrix computations through spatial-mode interference, utilizing inverse-designed multi-mode couplers, crossings, and Mach-Zehnder IQ modulators within a compact mode-division photonic tensor core (MPTC). Each guided mode (TE0-TE3) functions as a separate computational lane, facilitating matrix multiplications optically. This methodology overcomes the limitations and high costs associated with earlier PTAs, which depended on costly multi-wavelength light sources and large dot-product units featuring active phase-shifter components. This research is documented in a paper available on arXiv (2607.26016).

Key facts

  • MDTransformer is a hardware-software co-design of a photonic transformer accelerator.
  • It uses mode-division optical dataflow and operations.
  • Complex matrix operations are performed via spatial-mode interference.
  • Inverse-designed multi-mode couplers, crossings, and Mach-Zehnder IQ modulators are used.
  • A compact mode-division photonic tensor core (MPTC) is introduced.
  • Each guided mode (TE0-TE3) acts as an independent computational lane.
  • Matrix multiplications are executed in the optical domain.
  • Prior PTAs rely on expensive multi-wavelength light generation and large dot-product units.

Entities

Institutions

  • arXiv

Sources