Medium Access Control in Wireless Network-on-Chip: A Context Analysis

Sergi Abadal, Albert Mestres, Josep Torrellas, Eduard Alarcón, Albert Cabellos-Aparicio

TL;DR

As we move toward the "manycore era" with thousands of cores on a single chip, traditional wired connections are becoming a bottleneck. Wireless Network-on-Chip (WNoC) offers a high-speed alternative, but sharing the airwaves—Medium Access Control (MAC)—requires a total rethink. This paper provides a roadmap for designing these protocols by exploiting the unique, "monolithic" nature of computer chips.

What Problem Does This Paper Address?

In modern processor design, we are hitting a wall. Moving data between hundreds or thousands of cores using traditional copper wires and routers (Network-on-Chip or NoC) is slow and consumes too much power, especially for "broadcast" messages that need to reach every core at once.

Wireless communication on-chip (using millimeter-wave frequencies) is a promising fix, but it introduces a classic networking problem: how do you decide which core gets to "talk" on the wireless channel at any given time? Standard wireless protocols (like Wi-Fi or those for sensor networks) don't work here. They are designed for devices that move around, have batteries, and can tolerate milliseconds of delay. On a chip, cores are stationary, power is limited by heat, and a delay of even a few nanoseconds can tank performance.

Methods

The authors perform a context analysis, systematically breaking down the WNoC design space into three facets:

  1. Physical Constraints: Analyzing the "static and controlled landscape" of a silicon chip.
  2. Workload Characteristics: Examining how on-chip traffic is often bursty, has "hotspots" (certain cores talking more than others), and follows predictable patterns.
  3. Architectural Requirements: Defining the strict needs for ultra-low latency (1–100 ns) and extreme reliability (Bit Error Rates of $10^{-15}$).

The paper concludes by proposing a shift toward cross-layer design, where the MAC protocol isn't a standalone entity but is tightly integrated with the processor's compiler and the underlying wired network.

Datasets

While this is a position paper and does not introduce a new raw dataset, it synthesizes data from several key benchmarks and previous studies to characterize traffic:

  • SPLASH2 suite: Used to analyze the percentage of long-range and multicast traffic in cache-coherent systems.
  • fluidanimate application: Used to demonstrate "phase behavior," where a program alternates between intense communication and intense computation.
  • Traffic Models: The authors reference the Hurst exponent ($H$) and Gaussian standard deviation ($\sigma$) to quantify the "burstiness" and "spatial concentration" (hotspots) of typical manycore workloads.

Key Insights and Ideas

  • The "Monolithic" Advantage: Unlike a city-wide Wi-Fi network, a chip is a closed system. The designer knows exactly where every antenna is and can even use the compiler to predict when a core will need to send a burst of data. This allows for "proactive" MAC protocols rather than "reactive" ones.
  • Wired-Wireless Hybridization: WNoCs shouldn't replace wires; they should complement them. The authors suggest using the existing wired network as a "control plane" to handle synchronization and scheduling, leaving the wireless channel free for high-speed data.
  • Collision Detection is Back: In traditional wireless, a radio usually can't "hear" a collision while it is transmitting. However, because the on-chip environment is so stable and distances are so short (centimeters), the authors argue that we can implement streamlined collision detection—similar to early Ethernet—to drastically improve efficiency.
  • Energy vs. Latency: In WNoCs, you can't just turn the radio off to save power (as you would in a sensor network) because the latency hit is too high. Instead, protocols must "tune" the transceiver or use the wired network to wake up wireless interfaces just-in-time.

Key Results

The paper defines the "Gold Standard" requirements for a successful manycore WNoC MAC:

  • Latency: Must be between 1–100 ns.
  • Throughput: Needs to handle 10–100 Gb/s.
  • Reliability: A Bit Error Rate (BER) of $10^{-15}$, matching the reliability of physical wires.
  • Efficiency: Energy consumption should stay within 1–10 pJ/bit.

The authors highlight that while "token passing" (taking turns) is stable, it scales poorly. Conversely, "random access" (talking whenever) is fast but fails during traffic bursts. The "sweet spot" likely lies in hybrid protocols that adapt to the program being run.

Limitations and Future Work

The primary limitation is that many of these "compiler-assisted" and "predictive" MAC ideas are currently theoretical or in early-stage simulation. The authors note that:

  • Hardware Complexity: Creating hundreds of separate wireless channels (frequency multiplexing) is currently too hardware-intensive for a single chip.
  • ACK Implosion: When one core broadcasts to 1,000 others, the resulting flood of "Acknowledge" (ACK) messages can crash the network.
  • Future Work: The authors call for research into machine learning-based predictors that can sit on-chip and anticipate traffic patterns to reconfigure the MAC protocol in real-time.

Summary updated

Abstract

Wireless on-chip communication is a promising candidate to address the performance and efficiency issues that arise when scaling current NoC techniques to manycore processors. A WNoC can serve global and broadcast traffic with ultra-low latency even in thousand-core chips, thus acting as a natural complement to conventional and throughput-oriented wireline NoCs. However, the development of MAC strategies needed to efficiently share the wireless medium among the increasing number of cores remains a considerable challenge given the singularities of the environment and the novelty of the research area. In this position article, we present a context analysis describing the physical constraints, performance objectives, and traffic characteristics of the on-chip communication paradigm. We summarize the main differences with respect to traditional wireless scenarios, and then discuss their implications on the design of MAC protocols for manycore WNoC, with the ultimate goal of kickstarting this arguably unexplored research area.
BibTeX
@article{abadal2018medium,
  author = {Sergi Abadal and Albert Mestres and Josep Torrellas and Eduard Alarcón and Albert Cabellos-Aparicio},
  title = {{Medium Access Control in Wireless Network-on-Chip: A Context Analysis}},
  journal = {IEEE Communications Magazine},
  year = {2018},
  url = {https://openreview.net/forum?id=Upt8i17A34},
}

Loading related content...

Figure