TurnRL
RL-driven turn-restricted adaptive routing for chiplet-based systems (ICCAD 2026)
Advisors: Prof. Nathaniel Bleier and Prof. Ronald Dreslinski, University of Michigan · 2024/09 – 2025/11
- Developed fast turn-addition algorithms and baselines for arbitrary topologies and traffic patterns.
- Applied Deep Q-Learning with a GNN to learn optimal turn sets under different traffic patterns, evaluated with gem5 Garnet and PARSEC workloads.
- Improved saturation injection rate by 35.9% and 33.3% on HexaMesh and MI300A-like topologies, and reduced queuing and end-to-end latency by 20.1% and 8.8% on congested NAS-NPB workloads.
Publication: TurnRL: Reinforcement-Learning Driven Turn-Restricted Adaptive Routing for Chiplet-Based System (ICCAD 2026, first author).