Physical parameter-aware Networks-on-Chip design

Abstract

Networks-on-Chip (NoCs) have been proposed as a scalable, reliable and power-efficient communication fabric for chip multiprocessors (CMPs) and multiprocessor systems-on-chip (MPSoCs). NoCs determine both the performance and the reliability of such systems, with a significant power demand that is expected to increase due to developments in both technology and architecture. In terms of architecture, an important trend in many-core systems architecture is to increase the number of cores on a chip while reducing their individual complexity. This trend increases communication power relative to computation power. Moreover, technology-wise, power-hungry wires are dominating logic as power consumers as technology scales down. For these reasons, the design of future very large scale integration (VLSI) systems is moving from being computation-centric to communication-centric. On the other hand, chip’s physical parameters integrity, especially power and thermal integrity, is crucial for reliable VLSI systems. However, guaranteeing this integrity is becoming increasingly difficult with the higher scale of integration due to increased power density and operating frequencies that result in continuously increasing temperature and voltage drops in the chip. This is a challenge that may prevent further shrinking of devices. Thus, tackling the challenge of power and thermal integrity of future many-core systems at only one level of abstraction, the chip and package design for example, is no longer sufficient to ensure the integrity of physical parameters. New designtime and run-time strategies may need to work together at different levels of abstraction, such as package, application, network, to provide the required physical parameter integrity for these large systems. This necessitates strategies that work at the level of the on-chip network with its rising power budget. This thesis proposes models, techniques and architectures to improve power and thermal integrity of Network-on-Chip (NoC)-based many-core systems. The thesis is composed of two major parts: i) minimization and modelling of power supply variations to improve power integrity; and ii) dynamic thermal adaptation to improve thermal integrity. This thesis makes four major contributions. The first is a computational model of on-chip power supply variations in NoCs. The proposed model embeds a power delivery model, an NoC activity simulator and a power model. The model is verified with SPICE simulation and employed to analyse power supply variations in synthetic and real NoC workloads. Novel observations regarding power supply noise correlation with different traffic patterns and routing algorithms are found. The second is a new application mapping strategy aiming vii to minimize power supply noise in NoCs. This is achieved by defining a new metric, switching activity density, and employing a force-based objective function that results in minimizing switching density. Significant reductions in power supply noise (PSN) are achieved with a low energy penalty. This reduction in PSN also results in a better link timing accuracy. The third contribution is a new dynamic thermal-adaptive routing strategy to effectively diffuse heat from the NoC-based threedimensional (3D) CMPs, using a dynamic programming (DP)-based distributed control architecture. Moreover, a new approach for efficient extension of two-dimensional (2D) partially-adaptive routing algorithms to 3D is presented. This approach improves three-dimensional networkon- chip (3D NoC) routing adaptivity while ensuring deadlock-freeness. Finally, the proposed thermal-adaptive routing is implemented in field-programmable gate array (FPGA), and implementation challenges, for both thermal sensing and the dynamic control architecture are addressed. The proposed routing implementation is evaluated in terms of both functionality and performance. The methodologies and architectures proposed in this thesis open a new direction for improving the power and thermal integrity of future NoC-based 2D and 3D many-core architectures.