Answers are grounded in this paper and cite their sources.
Machine learning (ML) and artificial intelligence (AI) require more computing capabilities that have access to large memory and storage. Heterogeneous chiplet integrations are expected to break Moore's law of device scaling and achieve high-performance signal processing while increasing energy efficiency, in AI, graphics and other potential applications. The chiplet integration has been enabled by the development of advanced packaging technologies such as EMIB, CoWoS, Fan-out wafer level package (FOWLP). In this work the electrical characterization of a four-chiplets with four HBM3 modules system integrated using in FOWLP platform is discussed. The package uses a six metal layers RDL with dielectric polymer and 1.5 um width and space as minimum design rules. The Universal Chiplet Interconnect Express (UCIe) interface die-to-die standard is used to interconnect the four chiplets in a ring topology. The physical routing of the HBM3 interface requires about 1972 lines to be routed within available routing space approximately 8.8mm. Besides, the UCIe bus requires 156 lanes in routing space 388.8um between each two chiplets, are routed on the middle four layers of the FOWLP. The top and bottom two metal layers of the package are used to distribute the ground and power necessary for all four chiplets. This paper is to present the electrical performances on FOWLP by simulation analysis. And two possible (GSG/SSS) stack-up configurations were considered. Simulation showed that GSG layout design provide the best signal integrity to achieve specification of HBM3 and UCIe 1.1.
No takes yet. Share an insight, caveat, or question.
Zhuang et al. (2024) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: