Adaptive Gramian-Angular-Field segmentation integration based Generative Adversarial Network (AGSI-GAN) for eye diagram estimation of high bandwidth memory interposer

Abstract

With the advent of large-scale generative artificial intelligence, the number of I/Os and data rate for next-generation high bandwidth memory (HBM) to support supercomputing systems are rapidly increasing. Accordingly, signal and power integrity (SI/PI) analysis of the eye diagram, considering the design process of HBM interposer channel, is essential. In this paper, I propose an eye diagram estimation methodology using the adaptive Gramian-Angular-Field segmentation integration based generative adversarial network (AGSI-GAN) for rapid and accurate design. A conditional generative adversarial network (cGAN) for image-to-image transformation was implemented into a network suitable for eye diagram estimation. After setting modules reflecting SI/PI, I underwent the process of segmentation and integration, and extracted images through AGSI transformation. By providing the network with condition images that project memory channel characteristics, I enhanced the learning efficiency of the network. To validate the proposed methodology, I designed the hierarchical power distribution network (PDN) of HBM I/O interfaces and the channel of silicon interposers. Through this, I quickly and accurately estimated the eye diagram, including simultaneous switching noise (SSN) and far-end crosstalk (FEXT). As a result, I verified that the proposed method has high time efficiency and much higher estimation accuracy compared to the results of applying the conventional Gramian-Angular-Field (GAF) as a condition image.