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

Imagine you're building a system that needs to process vast amounts of data quickly and efficiently, like a real-time AI assistant. Current solutions often rely on electronic accelerators, which can be slow and consume a lot of energy, especially when handling complex tasks like those required by Transformer models. These systems typically use expensive components for light generation and large dot-product units, leading to inefficiencies and high costs — this is what's called the limitations of traditional electronic accelerators.

To tackle these issues, MDTransformer proposes a new approach that combines hardware and software design to optimize photonic transformer accelerators. By using mode-division optical dataflow, it performs matrix operations through spatial-mode interference, allowing for parallel processing without the constraints of spectral filtering. This means that MDTransformer can execute complex calculations more efficiently, achieving significant reductions in area, power, and energy consumption while maintaining comparable latency to existing systems. For builders, this means a more practical and scalable solution for deploying high-performance AI systems that can operate effectively in real-world applications.

MDTransformer introduces a novel hardware-software co-design that significantly improves photonic transformer accelerators.

The experimental results demonstrate substantial improvements in area, power, and energy savings compared to existing methods.