TL;DR
This paper introduces OIL, a comprehensive library of silicon nanophotonic devices designed to enable system-level planning and analysis of photonic networks-on-chip (NoC). By integrating within-core optical routing with traditional NoC packet routing onto a single optical layer, this work aims to significantly improve chip performance, power efficiency, and silicon utilization for future high-performance computing systems.
What Problem Does This Paper Address?
As semiconductor technology continues to scale down, traditional copper interconnects face fundamental limitations. These include bottlenecks in communication bandwidth, clock frequency, power dissipation, and increased crosstalk noise. To overcome these challenges and sustain Moore's Law, optical interconnects are emerging as a promising alternative. While optical interconnects have been successful in long-haul communication, their integration onto chips (on-chip) is still an active area of research. Existing approaches often focus on either device fabrication or network architecture, but a unified framework for designing and analyzing these complex systems is lacking. This paper addresses the need for a comprehensive design automation (CAD) framework that can systematically explore and optimize on-chip photonic interconnects, considering both device-level characteristics and system-level architectural trade-offs.
Methods
The core contribution of this paper is the development of OIL (Optical Interconnect Library). OIL is a parameterized library of silicon nanophotonic devices, including modulators, photodetectors, switches, couplers, buffers, and waveguides. These devices are characterized to provide essential parameters for performance analysis and design exploration.
The paper also proposes a new holistic photonic Networks-on-Chip (NoC) architecture. This architecture integrates two key functionalities onto a single optical layer:
- Global photonic routing network-on-chip: For efficient core-to-core communication, utilizing Time Division Multiplex (TDM), Wavelength Division Multiplex (WDM), or Space Division Multiplex (SDM) schemes.
- Within-core optical waveguide routing: For optimizing critical nets within processing units, aiming to improve timing and performance.
This integrated approach is supported by a CAD flow that allows for timing-driven mapping and routing, considering power, loss, and density constraints. The library and architecture are designed to facilitate quantitative analysis of power consumption, communication latency, and overall performance.
Datasets
Not discussed in the paper. The paper focuses on a design library and architectural proposal, with simulations and analyses based on device models and parameters derived from existing literature and simulation tools.
Key Insights and Ideas
- Unified Optical Layer: The most significant insight is the proposal to consolidate both within-core optical routing and core-to-core NoC packet routing onto a single, dedicated photonic silicon layer. This approach maximizes the utilization of the optical layer and offers a more integrated design flow.
- OIL as a Design Enabler: The OIL library provides the crucial parameterized device models needed to bridge the gap between device physics and system-level design. This allows for quantitative exploration of trade-offs, which is essential for making informed design decisions.
- Within-Core Optical Routing: Moving critical, timing-sensitive nets to the optical layer within cores offers a novel way to overcome electrical interconnect limitations, especially as clock frequencies increase.
- Holistic CAD Flow: The paper emphasizes the need for a holistic CAD flow that considers both electrical and photonic layers simultaneously, enabling optimal mapping of nets to the photonic layer for maximum performance gains.
- Scalability Potential: The proposed architecture and library are positioned to address the scalability challenges of future chip multi-processors by leveraging the inherent advantages of optical interconnects.
Key Results
The paper presents qualitative and quantitative analyses demonstrating the potential benefits of the proposed OIL library and photonic NoC architecture:
- Performance Improvement: Figure 15 shows a significant reduction in interconnect delays when using the proposed hybrid routing (electrical + optical) compared to purely electrical routing, indicating potential for higher clock frequencies and improved core performance.
- Interconnect Insertion Loss Analysis: Figure 16 provides insertion loss analysis for various photonic network routers, allowing designers to select components that meet specific loss budgets. The analysis highlights trade-offs between router complexity and loss.
- Scalability Discussion: Figure 17 illustrates how optical interconnects can help overcome the memory access bandwidth bottleneck that limits the scalability of traditional multi-core processors. The proposed architecture aims to achieve performance targets (curve B, D) that are significantly better than current trends (curve A, C).
Limitations and Future Work
- Fabrication Challenges: The authors acknowledge that while promising, on-chip WDM devices still face significant fabrication challenges.
- On-chip Buffering: Nanophotonic buffers are not yet widely used for on-chip applications due to the asynchronous nature of packet switching and the lack of sequential logic functions on the photonic NoC layer requiring buffering.
- CAD Optimization Complexity: Optimizing the mapping of electrical nets to the photonic layer involves complex CAD optimization problems under various constraints (power, loss, density).
- Future Work: The paper suggests further exploration of OIL's application to evaluate the proposed architecture in terms of performance improvement, power consumption, insertion loss, and performance scalability. Continued advancements in device fabrication and cost reduction for nanophotonic devices are also highlighted as crucial for future integration.
Summary updated