DxPTA: An Architecture Design Space Exploration with Optical Dataflow-guided Strategy for HW/SW Co-Design of Photonic Transformer Accelerators

Researchers introduce DxPTA, a design space exploration methodology for optimizing photonic transformer accelerators (PTAs) through hardware/software co-design. The approach automatically identifies optimal PTA architectures for AI models like DeiT and BERT while meeting area, power, energy, and latency constraints, achieving 15.2x faster design exploration than exhaustive methods.

DxPTA addresses a critical bottleneck in photonic computing: the lack of automated, constraint-aware design methodologies for transformer accelerators. While photonic processors promise significant energy efficiency and speed advantages over electronic systems, their manual architecture design has remained time-intensive and application-agnostic. This research bridges that gap by developing a systematic exploration framework that considers optical dataflow characteristics to guide architecture parameter selection.

The timing aligns with accelerating interest in alternative computing paradigms for AI workloads. As transformer models grow larger and energy demands become critical for data centers and edge devices, photonic solutions offer compelling advantages. However, the field has matured slowly partly due to design complexity—researchers previously relied on ad-hoc approaches without comprehensive constraint optimization. DxPTA's methodology demonstrates that algorithmic approaches can systematically explore the architecture design space while respecting practical deployment requirements.

For the AI hardware industry, this work signals maturation of photonic accelerator research. The methodology enables faster iteration cycles for different applications, potentially accelerating commercialization timelines. Achieving configurations within specified constraints (26mm² area, 4.8W power) for modern transformers validates the feasibility of photonic solutions for real deployments. This could reshape how semiconductor and AI infrastructure companies approach next-generation accelerator design, especially as power efficiency becomes a primary competitive factor.

The research establishes foundations for broader adoption of photonic computing in production environments. Future work will likely extend these methodologies to other neural network architectures and explore hybrid photonic-electronic systems that leverage strengths of both paradigms.

• →DxPTA automates photonic transformer accelerator design while meeting practical constraints like area, power, and latency requirements.
• →The methodology achieves 15.2x faster design exploration compared to exhaustive search approaches, enabling scalable architecture optimization.
• →Results demonstrate feasible implementations for BERT and DeiT models, validating photonic accelerators for real-world transformer deployment.
• →The approach leverages coherent optical dataflow analysis to systematically identify and evaluate architecture parameters.
• →This research accelerates photonic computing commercialization by reducing design iteration time and enabling constraint-aware optimization.