DesignCon 2022 Schedule Builder

As the data rate of PCIe dramatically increases, channel loss has reached the limit. To overcome the channel loss problems of PCIe 6.0., PAM-4 signaling is required to increase the bandwidth. Since the eye-margin of PAM-4 is 1/3 smaller compared to NRZ, controlling the PAM-4 based signaling system is extremely sensitive to parameters of high-speed serial links; the possibility of PAM-4 degrading SI than NRZ arises. Therefore, providing such an effective optimization method is crucial for the improvement of SI characteristics of PAM-4 based signaling system.

In this paper, we propose a novel BO scheme based on DNN objective-regressor for PAM-4 based PCIe 6.0. Our scheme has two main components: (a) BO and (b) DNN-based objective simulator. The BO learns to generate a candidate solution based on data collected from the objective simulator. To construct a fast, accurate and flexible objective simulator, we combine the circuit simulator and DNN model. To be specific, the circuit simulator produces S-parameters from the design solution provided by BO. Then the DNN returns eye-opening from the provided partial information of S-parameter. Since the pre-trained DNN model enables a quick computation of eye-opening, the proposed scheme can overcome the previous limitations of BO scheme.

Takeaway

This paper a novel BO scheme based on deep neural network (DNN) objective-regressor for PAM-4 based PCIe 6.0. The proposed method tunes whole system of PCIe 6.0 including high-speed channel and via in extremely fast speed. The proposed method can be also applied to various system-level interconnection optimization tasks.

Intended Audience

This paper a novel BO scheme based on deep neural network (DNN) objective-regressor for PAM-4 based PCIe 6.0. The proposed method tunes whole system of PCIe 6.0 including high-speed channel and via in extremely fast speed. The proposed method can be also applied to various system-level interconnection optimization tasks.