Summary: | 碩士 === 國立交通大學 === 電子研究所 === 106 === At advanced technology nodes, there exists a variety of power/ground network
(P/G network) structures. Due to the various structures, finding a robust
P/G becomes a time-consuming task. A robust P/G means a P/G which satisfies
the IR-drop constraint and occupies small routing resource. In order to shorten
the time of finding a robust P/G, we improve the existing industrial flow by using
machine-learning (ML) techniques. At first, we intend to insert an ML-predictor in
the industrial flow to narrow down the search space by removing failed P/Gs which
do not satisfy the IR-drop constraint. However, because of training data issues,
in this thesis, we propose a producing guidance data method to collect the proper
training data and use early stop to limit the number of produced training data.
Based on our techniques, we overcome the training data issues and expedite the
industrial flow. Experimental results show that we can reduce the runtime up to
48%. Furthermore, by analyzing the correlation between IR-drop and P/G structures,
we find that IR-drop is sensitive to the structure of middle and bottom layers
for advanced processes.
|