Cooling codes: Thermal-management coding for high-performance interconnects

Yeow Meng Chee, Tuvi Etzion, Han Mao Kiah, Alexander Vardy

Introduction

As semiconductor technology scales to 65nm and beyond, thermal management has emerged as a primary design constraint. High temperatures in on-chip interconnects (buses) lead to increased signal delay (Elmore delay) and exponential growth in leakage current, creating a vicious cycle of performance degradation and power consumption. While traditional "low-power" coding techniques focus on reducing average transition activity to save energy, they often fail to address localized "hot spots"—specific wires that reach dangerous temperatures even if the overall bus power is low.

The Thermal Challenge

To effectively manage the thermal health of a bus, designers must control both the temporal and spatial distribution of signal transitions. A few wires switching frequently can become significantly hotter than their neighbors.

Cross-sectional view of on-chip wires illustrating heat flow to the substrate layer.

As shown in the physical model above, heat generated in the wires must dissipate through the interlayer dielectric (ILD) to the substrate. The thermal resistance of these paths determines the steady-state temperature. "Cooling Codes" address this by providing a mechanism to "turn off" transitions on the hottest wires, allowing them to cool down while maintaining data throughput.

System Architecture and Coding Framework

The proposed solution integrates specialized encoder and decoder blocks at the ends of the Deep Sub-Micron (DSM) bus.

High-level architecture of a Deep Sub-Micron (DSM) bus with integrated cooling code encoder/decoder blocks.

The framework utilizes Differential Encoding, where a '1' in a codeword signifies a state transition (bit flip) and a '0' signifies no change. This direct mapping allows the system to enforce specific thermal properties:

  • Property A(t): Cooling the $t$ hottest wires by ensuring they remain static during transmission.
  • Property B(w): Maintaining low power by limiting the total transitions across all wires to $\le w$.
  • Property C(e): Providing error correction for up to $e$ transmission errors.

Mathematical Foundations

Constructing these codes requires sophisticated combinatorial designs. The authors leverage Projective Geometry (specifically spreads and partial spreads) to create optimal cooling codes. They also employ Steiner Systems and Baranyai’s Theorem to develop "equireplicate" codes that distribute transitions evenly over time, preventing any single wire from becoming a chronic hot spot. For scenarios requiring the simultaneous cooling of many wires, Sunflower Constructions from extremal set theory are used to maintain efficiency.

Key Insights

  • The $t+1$ Rule: A significant finding is that cooling $t$ specific wires only requires $t+1$ redundant wires ($n = k + t + 1$). This theoretical minimum makes the overhead for thermal management remarkably low, especially for wide buses.
  • Adaptive Control: Unlike static low-power codes, Cooling Codes are inherently adaptive. The encoder uses real-time data from thermal sensors or analytical counters to identify the hottest wires and select codewords that avoid them.
  • Asymptotic Optimality: As bus width ($k$) increases, the relative area overhead of these codes becomes negligible. The authors identified a threshold ($\tau \approx 0.687$); below this ratio of cooled wires to total wires, their constructions achieve asymptotic optimality.

Results and Future Work

The research proves that $n = k + t + 1$ is the optimal redundancy for cooling $t$ wires. Despite the underlying mathematical complexity, the resulting encoders and decoders are practical for hardware implementation, featuring linear complexity relative to bus width.

While the bounds for cooling are tight, future work remains in closing the gap for codes that combine cooling with strict low-power constraints. Additionally, while the theoretical area overhead is low, further research is needed to evaluate the practical overhead of integrating real-time thermal sensors and the feedback loop into the encoding logic.

Summary updated

BibTeX
@article{chee2018cooling,
  author = {Yeow Meng Chee and Tuvi Etzion and Han Mao Kiah and Alexander Vardy},
  title = {{Cooling codes: Thermal-management coding for high-performance interconnects}},
  journal = {IEEE Trans. Information Theory},
  year = {2018},
  doi = {10.1109/tit.2017.2771245},
}
Figure