TTQR: A Traffic- and Thermal-Aware Q-Routing for 3D Network-on-Chip.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Thermal Management

Background:

• 3D network-on-chips (3D NoCs) face reliability and performance issues due to high power density and thermal gradients.
• Unequal thermal conductance in stacked layers exacerbates these problems, leading to degradation.
• Congestion-aware adaptive routing can mitigate these issues by balancing network traffic load.

Purpose of the Study:

• To propose a novel routing algorithm for 3D NoCs that addresses both traffic congestion and thermal issues.
• To enhance the reliability and performance of 3D NoCs through intelligent routing decisions.
• To reduce hardware overhead while improving network efficiency.

Main Methods:

• Developed a traffic- and thermal-aware Q-routing (TTQR) algorithm based on Q-learning, a reinforcement learning technique.
• Utilized two Q-tables (Q1 for traffic, Q2 for temperature) updated via packet headers, minimizing hardware requirements.
• Packet output port selection is determined by the ratio of Q1 to Q2 values, directing traffic to optimize thermal conditions and inter-layer balance.

Main Results:

• The TTQR algorithm effectively alleviates thermal problems and balances inter-layer traffic.
• Simulations using the Noxim platform demonstrated significant performance improvements over the TAAR routing algorithm.
• Compared to TAAR, TTQR achieved an average reduction in latency of 63.6% and an average increase in throughput of 41.4% for synthetic traffic patterns.

Conclusions:

• The proposed TTQR algorithm offers a robust solution for enhancing 3D NoC performance and reliability.
• Q-learning provides an effective framework for developing adaptive routing strategies that consider both traffic and thermal dynamics.
• TTQR presents a hardware-efficient approach to overcoming the challenges posed by 3D integration in network-on-chips.