DesignCon 2022 Schedule Builder

This paper proposes a practical and reusable decoupling capacitor (decap) placement solver by imitation learning (IL). The proposed IL framework imitates an expert policy and learns the policy that guarantees performance beyond that of expert policy and reusability in terms of PDN with different probing port and keep-out region; the constructed policy itself becomes the solver. The expert policy can be any algorithm or conventional tool, which means this is a fast and effective approach to improve existing methods. Previous works on machine learning based decap placement optimization methods seems fancy but are not practical; none of them are used in industry. They have only shown the feasibility of applying machine learning, not shown any quantitative analysis of their performance in comparison to other methods. The proposed method, on the other hand, provides detailed comparison to other methods. Moreover, by taking a conventional tool used in industry as an expert policy, the proposed method can construct a reusable decap placement solver that is data-efficient, practical and guarantees better performance than the conventional tool. In this paper, genetic algorithm is taken as an expert policy to verify how the proposed method generates a solver that learns beyond the level of the expert policy, and time-performance analysis of various methods are presented.

Takeaway

This paper proposes a practical and reusable decap placement solver by imitation learning (IL). The proposed IL framework imitates an expert policy and learns the policy that guarantees performance beyond that of expert policy and reusability in terms of PDN with different probing port and keep-out region; expert policy can be any algorithm or conventional tool and the constructed policy itself becomes the solver.

Intended Audience

This paper proposes a practical and reusable decoupling capacitor placement solver by imitation learning (IL). The proposed IL framework imitates an expert policy and learns the policy that guarantees performance beyond that of expert policy and reusability in terms of PDN with different probing port and keep-out region; expert policy can be any algorithm or conventional tool and the constructed policy itself becomes the solver.