Abstract

This talk presents new ideas for extremely complex geometries in Integrated Systems design with special emphasis on developments needed for AI applications. Efficiency of computations are among most fundamental requirements for dealing with massive amounts of data and power and time limitations. Classical IC design relies on so-called “planar circuit” concept proposed by Jean Hernie from Fairchild Semiconductors. Since those early days this principle remains still  a fundament for many kninds of mocro-nano system integration. Extreme requirements posed by new applications require new approachces – building circuits in 3D is the first idea already brought to industrial products. Next the “chiplet” concept was brought to standard operation and is widely considered by leading chip manufacturers. Further power reduction, parallelization of computations and massive data transfer will require new groundbreaking solutions. What if the “planar circuit” assumption is dropped? and the buiding blocks and subcircuits positioning has no constraints. In such a system design not only placement is crucial- routing ie. Making all necessary types of connections has to break the limit of Steiner-type tree structures. We present some new concepts in the design and also new solutions methodologies for such design based on Machine Learning.