Cooling Codes: Thermal-Management Coding for High-Performance Interconnects

TL;DR

This paper introduces innovative coding schemes called cooling codes that directly manage the thermal profile of high-performance interconnects by controlling the hottest wires, reducing peak temperature and power consumption while enabling error correction.

Contribution

The paper presents the first coding schemes that simultaneously control peak temperature, reduce average power, and support error correction for high-performance interconnects.

Findings

Cooling codes effectively lower peak temperature of hottest wires.

The minimum number of wires needed is n = k + t + 1 for cooling t hottest wires.

Proposed schemes are practical with low computational overhead.

Abstract

High temperatures have dramatic negative effects on interconnect performance and, hence, numerous techniques have been proposed to reduce the power consumption of on-chip buses. However, existing methods fall short of fully addressing the thermal challenges posed by high-performance interconnects. In this paper, we introduce new efficient coding schemes that make it possible to directly control the peak temperature of a bus by effectively cooling its hottest wires. This is achieved by avoiding state transitions on the hottest wires for as long as necessary until their temperature drops off. We also reduce the average power consumption by making sure that the total number of state transitions on all the wires is below a prescribed threshold. We show how each of these two features can be coded for separately or, alternatively, how both can be achieved at the same time. In addition, error-correction for the transmitted information can be provided while controlling the peak temperature and/or the average power consumption. In general, our cooling codes use $n > k$ wires to encode a given $k$-bit bus. One of our goals herein is to determine the minimum possible number of wires $n$ needed to encode $k$ bits while satisfying any combination of the three desired properties. We provide full theoretical analysis in each case. In particular, we show that $n = k+t+1$ suffices to cool the $t$ hottest wires, and this is the best possible. Moreover, although the proposed coding schemes make use of sophisticated tools from combinatorics, discrete geometry, linear algebra, and coding theory, the resulting encoders and decoders are fully practical. They do not require significant computational overhead and can be implemented without sacrificing a large circuit area.