Network-on-chip architectures and design methods

Luca Benini, Davide BertozziIEE 2005 · DOI

TL;DR

As System-on-Chips (SoCs) grow in complexity, traditional bus-based communication architectures are hitting scalability limits. This paper surveys the evolution from these buses to Networks-on-Chip (NoCs), highlighting how NoCs offer a more scalable and robust solution for on-chip communication by adopting principles from macroscopic networks.

What Problem Does This Paper Address?

The relentless scaling of semiconductor technology has led to Systems-on-Chip (SoCs) with an ever-increasing number of integrated cores. This massive integration demands sophisticated communication infrastructure to connect these cores efficiently. However, traditional on-chip bus architectures, while simple and area-efficient, struggle to scale with the growing number of cores and the increasing communication demands. This leads to performance bottlenecks, power inefficiencies, and design complexity issues. The paper addresses this challenge by exploring the evolution of on-chip communication from advanced bus protocols to the more revolutionary Network-on-Chip (NoC) paradigm, aiming to identify guiding principles for future designs.

Methods

The paper presents a survey of on-chip communication architectures, tracing their evolution. It starts by discussing the limitations of traditional bus architectures, particularly in terms of scalability and modularity. It then delves into the advancements in bus specifications, using the AMBA (Advanced Micro-Controller Bus Architecture) as a case study. This includes examining multilayer AHB (Advanced High-performance Bus) for improved bandwidth through crossbar interconnects and AMBA AXI (Advanced eXtensible Interface) for its point-to-point, pipelined, and multi-channel approach. Finally, the paper introduces Networks-on-Chip (NoCs) as a more radical solution, discussing their fundamental architecture, including links, switches, and network interfaces, and exploring design methodologies and early prototypes. The analysis involves quantitative evaluation through simulations and performance comparisons of different architectures under various traffic loads.

Datasets

Not discussed in the paper. The paper focuses on architectural survey and performance analysis using simulation models and theoretical evaluations rather than training on specific datasets.

Key Insights and Ideas

  • The "Interconnect Showstopper": As SoCs scale, the communication fabric becomes the primary bottleneck, limiting performance and power efficiency.
  • Bus Evolution: Buses have evolved significantly (e.g., AMBA AHB to AXI) to improve parallelism, bandwidth, and modularity, but fundamental scalability limitations remain.
  • NoCs as a Paradigm Shift: Networks-on-Chip adopt principles from macroscopic networks (packet switching, routing, flow control) to provide a inherently scalable and robust communication infrastructure.
  • Point-to-Point (PP) Interfaces are Key: Decoupling cores from the interconnect through PP interfaces simplifies design and enhances modularity, a trend seen in AXI and fundamental to NoCs.
  • Topology Matters: The choice of interconnect topology (e.g., shared bus, crossbar, mesh, custom) has a profound impact on performance, area, and power, especially under heavy traffic.
  • Application-Specific NoCs: For complex SoCs with heterogeneous cores and communication patterns, custom NoC designs can offer significant advantages in performance, power, and area compared to regular topologies.

Key Results

  • AMBA AXI vs. AHB: Under heavy traffic loads with small caches (causing frequent misses), AMBA AXI demonstrated significantly better scalability than AHB, with execution times increasing by only 12-15% compared to 77% for AHB when moving from 2 to 8 cores. AXI achieved transfer efficiencies of up to 81% compared to 47% for AHB.
  • Topology Impact: In a matrix multiplication benchmark, a multilayer (crossbar) topology achieved a speedup of over two times compared to a shared bus. A bridged bus offered a lower speedup, highlighting the sensitivity to traffic patterns.
  • Custom NoC Advantages: For an MPEG4 decoder application, custom NoC configurations (COSTUM 1 and CUSTOM 2) showed reduced area (up to 1.85x smaller than a mesh) and competitive power consumption compared to a standard mesh topology. Custom NoCs also demonstrated lower average packet latency.
  • NoC Scalability: NoCs are presented as the solution to overcome the fundamental scalability limitations of single-hop interconnects, enabling multihop topologies and efficient routing.

Limitations and Future Work

  • Complexity and Area: Advanced bus architectures and NoCs, while offering performance benefits, come at the cost of increased silicon area and design complexity. Crossbar interconnects, for instance, become unmanageable for a large number of masters and slaves.
  • Design Methodology Gap: The evolution towards NoCs necessitates new design skills and methodologies, particularly for mapping complex applications onto heterogeneous NoC architectures and for automating the synthesis of custom NoCs.
  • Custom NoC Synthesis: While custom NoCs offer advantages, their design and synthesis are complex. Tools like XpipesCompiler are emerging to automate this, but the quality of synthesized components might not match highly optimized hard macros. Further work is needed in backend synthesis (RTL, placement, routing) to fully realize the benefits.
  • Synchronisation: Supporting flexible synchronisation schemes (like GALS - Globally Asynchronous, Locally Synchronous) is crucial for large-scale SoCs, as distributing a single high-frequency clock becomes unmanageable.
  • Application-Specific Mapping: The effectiveness of NoCs is highly dependent on how well the application's communication requirements are mapped onto the network topology. This remains a critical design challenge.

Summary updated

Abstract

Performance and power of gigascale systems-on-chip (SoCs) is increasingly communication-dominated. Designers have to accommodate the communication needs of an increasing number of integrated cores while preserving overall system performance under tight power budgets. State-of-the-art SoC communication architectures start facing scalability as well as modularity limitations, and more advanced bus specifications are emerging to deal with these issues at the expense of silicon area and complexity. Communication architecture evolutions mainly regard bus protocols (to better exploit available bandwidth) and bus topologies (to increase bandwidth). In the long run, more aggressive solutions are needed to overcome the scalability limitation, and networks-on-chip (NoCs) are currently viewed as a ‘revolutionary’ approach to provide a scalable, high performance and robust infrastructure for on-chip communication. The paper aims at surveying the evolution of the field, moving from SoC buses to forward-looking NoC research prototypes. The elements of continuity, as well as the key differences, will be captured, in an effort to extract general guiding principles in a fast-evolving domain.
BibTeX
@inproceedings{benini2005network,
  author = {Luca Benini and Davide Bertozzi},
  title = {{Network-on-chip architectures and design methods}},
  booktitle = {IEE},
  year = {2005},
  doi = {10.1049/ip-cdt:20045100},
}
Figure