In this paper, we propose a novel channel routing scheme, which automatically designs effective routing solvers without heuristic knowledge. Our method is based on two-stage components, the deep reinforcement learning (DRL) framework for the automation of solver design and Bayesian optimization for fine-tuning the channels routed by the designed solver. The agent of the DRL framework is parametrized by the Transformer, a cutting-edge language model applied to BERT and GPT-3. We represent the channel routing problem as a sequential decision processing, and we take advantage of the powerful sequential processing capabilities of the language model (Transformer).
The trained Transformer becomes a channel routing solver, termed the initial router, which finishes pin-to-pin routing roughly considering signal integrity (SI). Then, the proposed Bayesian optimization scheme, termed the post router fine-tunes the physical parameters of the channel routed by the initial router, considering SI.
Extensive experiments demonstrate that the proposed routing solvers outperform baseline routing algorithms in several test cases, including routing on high bandwidth memory (HBM) interposer, at a significantly faster speed.