MDTransformer · Jul 28, 2026

MDTransformer: A Hardware-Software Co-Design of Mode-Division Photonic Transformer Accelerator with Inverse-Designed Coherent Crossbar

MDTransformer is a photonic transformer accelerator that uses mode-division multiplexing with TE0–TE3 guided modes as independent computational lanes, achieving four-fold parallelism per waveguide without spectral filtering. It employs inverse-designed multi-mode couplers, crossings, and Mach-Zehnder IQ modulators in a mode-division photonic tensor core (MPTC) for complex-valued matrix multiplications via coherent detection and IQ modulation.

What happened

MDTransformer introduces a hardware-software co-design for photonic transformer accelerators, replacing expensive multi-wavelength approaches with mode-division optical dataflow. Its MPTC leverages spatial-mode interference and inverse-designed components to perform complex matrix operations, enabling compact, efficient Transformer inference acceleration.

Technical significance

The design uses spatial-mode interference and inverse-designed photonic components to create a compact tensor core that performs complex-valued matrix multiplications without wavelength-division multiplexing, potentially reducing hardware complexity and cost.

Industry impact

This approach could lower the barrier to photonic AI accelerators by eliminating the need for multi-wavelength lasers and large dot-product units, making photonic inference more practical for data centers and edge deployments.

What to watch

Next signals include experimental validation of the MPTC, benchmarking against electronic accelerators on standard Transformer models, and exploration of scalability to more modes or integration with electronic control systems.

Decision value

If realized, MDTransformer could offer significant energy efficiency and throughput gains for Transformer inference, reducing operational costs for AI service providers and enabling new applications in power-constrained environments.

Evidence