Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Large-Scale Photonics-Empowered AI Systems: From Physical Design Automation to System-Algorithm Co-Exploration

Ziang Yin, Hongjian Zhou, Nicholas Gangi, Meng Zhang, Jeff Zhang, Zhaoran Rena Huang, Jiaqi Gu

TL;DR

The paper tackles the challenge of scaling photonics-empowered AI by proposing a cross-layer paradigm that couples device physics, circuit design, architecture, and learning algorithms through automated EPDA and system–algorithm co-exploration. It introduces SimPhony for physics-aware system modeling, ADEPT and ADEPT-Z for automated topology exploration, and Apollo & LiDAR for automated EPDA-based layout and routing, all demonstrated alongside Lightening-Transformer, TeMPO, and SCATTER tensor-core designs. The work reports substantial gains in latency, energy efficiency, and area reduction, and addresses robustness to non-idealities and layout constraints, illustrating a feasible path to manufacturable, large-scale photonic AI systems. By tightly integrating physical feasibility with architectural and algorithmic decisions, this framework enables scalable exploration of photonic fabrics and practical deployment at cloud-to-edge scales.

Abstract

In this work, we identify three considerations that are essential for realizing practical photonic AI systems at scale: (1) dynamic tensor operation support for modern models rather than only weight-static kernels, especially for attention/Transformer-style workloads; (2) systematic management of conversion, control, and data-movement overheads, where multiplexing and dataflow must amortize electronic costs instead of letting ADC/DAC and I/O dominate; and (3) robustness under hardware non-idealities that become more severe as integration density grows. To study these coupled tradeoffs quantitatively, and to ensure they remain meaningful under real implementation constraints, we build a cross-layer toolchain that supports photonic AI design from early exploration to physical realization. SimPhony provides implementation-aware modeling and rapid cross-layer evaluation, translating physical costs into system-level metrics so architectural decisions are grounded in realistic assumptions. ADEPT and ADEPT-Z enable end-to-end circuit and topology exploration, connecting system objectives to feasible photonic fabrics under practical device and circuit constraints. Finally, Apollo and LiDAR provide scalable photonic physical design automation, turning candidate circuits into manufacturable layouts while accounting for routing, thermal, and crosstalk constraints.

Toward Large-Scale Photonics-Empowered AI Systems: From Physical Design Automation to System-Algorithm Co-Exploration

TL;DR

The paper tackles the challenge of scaling photonics-empowered AI by proposing a cross-layer paradigm that couples device physics, circuit design, architecture, and learning algorithms through automated EPDA and system–algorithm co-exploration. It introduces SimPhony for physics-aware system modeling, ADEPT and ADEPT-Z for automated topology exploration, and Apollo & LiDAR for automated EPDA-based layout and routing, all demonstrated alongside Lightening-Transformer, TeMPO, and SCATTER tensor-core designs. The work reports substantial gains in latency, energy efficiency, and area reduction, and addresses robustness to non-idealities and layout constraints, illustrating a feasible path to manufacturable, large-scale photonic AI systems. By tightly integrating physical feasibility with architectural and algorithmic decisions, this framework enables scalable exploration of photonic fabrics and practical deployment at cloud-to-edge scales.

Abstract

In this work, we identify three considerations that are essential for realizing practical photonic AI systems at scale: (1) dynamic tensor operation support for modern models rather than only weight-static kernels, especially for attention/Transformer-style workloads; (2) systematic management of conversion, control, and data-movement overheads, where multiplexing and dataflow must amortize electronic costs instead of letting ADC/DAC and I/O dominate; and (3) robustness under hardware non-idealities that become more severe as integration density grows. To study these coupled tradeoffs quantitatively, and to ensure they remain meaningful under real implementation constraints, we build a cross-layer toolchain that supports photonic AI design from early exploration to physical realization. SimPhony provides implementation-aware modeling and rapid cross-layer evaluation, translating physical costs into system-level metrics so architectural decisions are grounded in realistic assumptions. ADEPT and ADEPT-Z enable end-to-end circuit and topology exploration, connecting system objectives to feasible photonic fabrics under practical device and circuit constraints. Finally, Apollo and LiDAR provide scalable photonic physical design automation, turning candidate circuits into manufacturable layouts while accounting for routing, thermal, and crosstalk constraints.
Paper Structure (10 sections, 2 figures)

This paper contains 10 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Photonics-Empowered AI Systems: A Cross-Layer View
  3. Lightening-Transformer: Dynamic Tensor Operations for Transformer Inference
  4. TeMPO: Amortizing Conversion Overheads for Edge-Efficient Photonic AI
  5. SCATTER: Robust and Scalable Photonic Tensor Cores under Physical Non-Idealities
  6. Electronic-Photonic Design Automation: Key Enabler of Scalable Photonic AI Systems
  7. SimPhony: Cross-Layer Modeling from Device Response to System Performance
  8. ADEPT: Automated Multi-Objective PIC Topology Exploration
  9. Apollo & LiDAR: Automated EPDA Flow for PICs and EPICs
  10. Discussion

Figures (2)

  • Figure 1: Overview of the SimPhony cross-layer modeling and co-exploration framework. Device- and circuit-level photonic models are integrated with architectural analysis, dataflow mapping, and power, area, and memory estimation. Hardware-aware training and conversion are supported through tight coupling with learning frameworks, enabling system--algorithm co-exploration under realistic physical constraints.
  • Figure 2: Our proposed automated PIC placement engine Apollo apollo and router LiDAR-V2 LiDAR_ISPD_ZhouLiDAR2_ARXIV_Zhou can generate compact, high-quality layout for large-scale PICs (over 1000 devices) within 230s.