Low-power integration of on-chip nanophotonic interconnect for high-performance optoelectrical IC

Duo Ding, David Z. PanProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2012 · DOI

TL;DR

This paper introduces a novel design automation framework, O-Router and GLOW, for creating low-power, high-performance on-chip nanophotonic interconnects. By leveraging flexible, non-rectilinear routing and integrating Wavelength Division Multiplexing (WDM), the authors demonstrate significant power reductions and improved performance compared to traditional electrical interconnects, paving the way for more efficient application-specific integrated circuits.

What Problem Does This Paper Address?

As semiconductor technology pushes towards higher integration densities and performance, traditional electrical interconnects are becoming a bottleneck. The increasing ratio of interconnect delay to gate delay, coupled with the complexity of timing closure in large designs, makes electrical interconnects power-hungry and difficult to manage. Nanophotonics, with its inherent speed and low power potential, offers a promising alternative. However, designing and integrating these complex photonic components, which have unique physical constraints (like thermal sensitivity and specific device requirements), into a functional system is a significant challenge. This paper addresses the need for automated design methodologies that can efficiently explore the design space of nanophotonic interconnects, optimize their placement and routing, and ensure they meet critical performance and reliability requirements.

Methods

The authors propose a two-pronged approach:

  • O-Router: This system focuses on optimizing the physical layout of nanophotonic interconnects. It moves beyond traditional Manhattan (rectilinear) routing by exploring flexible, non-rectilinear geometries. O-Router uses an Integer Linear Programming (ILP) formulation to find optimal routing paths that minimize total modulating power while satisfying various device constraints, such as signal integrity, throughput, and power thresholds. It considers different types of optical losses, including straight-line loss, bending loss, and coupling loss, to make informed routing decisions.
  • GLOW: This is a more comprehensive global routing framework that integrates Wavelength Division Multiplexing (WDM) for high-density interconnects. GLOW addresses thermal-aware, low-power physical synthesis by considering WDM waveguide placement, optical channel allocation, and optical-electrical data converter planning. It employs an ILP-based approach to optimize for power efficiency and thermal reliability, taking into account factors like WDM trunk crossings, thermal effects on ring resonators, and propagation losses. GLOW also incorporates a netlist pre-processing step to derive optical netlists from electrical benchmarks and a hierarchical clustering method to group pins, reducing the number of necessary opto-electrical conversions.

Datasets

The paper utilizes benchmarks derived from ISPD (International Symposium on Physical Design) global routing contest benchmarks. These benchmarks, originally containing over 100K nets, are scaled up to a centimeter-scale for the purpose of this study. The authors refer to these scaled-up optical netlists as CK1-CK6. They also mention using standard predictive technology models (PTM) for MOSFET models.

Key Insights and Ideas

  • Non-rectilinear routing is crucial for nanophotonics: Unlike electrical interconnects where sharp turns are penalized, optical waveguides can be routed with flexible geometries, potentially reducing overall loss and improving efficiency.
  • WDM is key to high-density interconnects: By multiplexing multiple signals onto a single waveguide using different wavelengths, WDM significantly increases the capacity of on-chip optical links.
  • Thermal reliability is a major concern: Nanophotonic devices, particularly ring resonators, are sensitive to temperature variations, which can cause wavelength drift and performance degradation. The proposed methods explicitly account for these thermal constraints.
  • Automated design is essential: The complexity of nanophotonic devices and their interconnections necessitates sophisticated CAD tools to explore the vast design space and find optimal solutions that satisfy multiple, often conflicting, constraints.
  • ILP provides a powerful optimization framework: The use of Integer Linear Programming allows for the formulation of complex objectives and constraints, enabling the optimization of global routing for power and thermal performance.

Key Results

The paper demonstrates significant improvements in power consumption and performance:

  • O-Router: Achieved power reductions of 22%-49% compared to a Minimum Spanning Tree (MST) routing algorithm on various benchmarks, with a maximum reduction of 81.1% on one case.
  • GLOW: Showed power reductions of 22%-49% compared to a greedy heuristic approach (CAT) on the same benchmarks, with a maximum reduction of 67.3% on one case.
  • GLOW's approach, while potentially leading to slightly longer waveguide lengths, prioritizes global power optimization by strategically assigning WDM trunks and channels, effectively minimizing overall power consumption.

Limitations and Future Work

  • ILP solver limitations: In cases where no feasible solutions exist, the ILP formulation may require adjustments to the initial WDM trunk placement. The authors mention a progressive and heuristic approach to address this.
  • Scalability of ILP: While the authors state that accelerated ILP manages to find optimal solutions within acceptable runtimes for the tested optical routing problem sizes, the scalability of ILP for extremely large and complex netlists might still be a concern.
  • Future Work: The authors suggest further research in co-optimizing CAD and nanophotonics technologies using automated design techniques. This implies exploring tighter integration between the physical design tools and the underlying fabrication processes and device designs.

Summary updated

Abstract

In this manuscript we study the potentials of nanophotonics on-chip integration and propose a set of automation methodologies to construct low power on-chip interconnect with flexible geometry shapes. We show that with such techniques, a systematic design aid environment can be developed to generate optimized integration configurations meanwhile honoring complex sets of photonic device constraints. Due to their unique characteristics, not only do these techniques benefit the optimization of on-chip photonic networks, but also they can be efficiently applied to build low-power high-throughput application specific ICs with opto-electrical interconnection.
BibTeX
@inproceedings{ding2012low,
  author = {Duo Ding and David Z. Pan},
  title = {{Low-power integration of on-chip nanophotonic interconnect for high-performance optoelectrical IC}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2012},
  doi = {10.1117/12.913886},
}
Figure