Performance-Driven X-Architecture Routing Algorithm for Artificial Intelligence Chip Design in Smart Manufacturing

Abstract

The new 7-nm Artificial Intelligence (AI) chip is an important milestone recently announced by the IBM research team, with a very important optimization goal of performance. This chip technology can be extended to various business scenarios in the Internet of Things. As the basic model for very large scale integration routing, the Steiner minimal tree can be used in various practical problems, such as wirelength optimization and timing closure. Further considering the X-architecture and the routing resources within obstacles, an effective performance-driven X-architecture routing algorithm for AI chip design in smart manufacturing is proposed to improve the delay performance of the chip. First, a special particle swarm optimization algorithm is presented to solve the discrete length-restricted X-architecture Steiner minimum tree problem in combination with genetic operations, and a particle encoding scheme is presented to encode each particle into an initial routing tree. Second, two lookup tables based on pins and obstacles are established to provide a fast information query for the whole algorithm flow. Third, a strategy of candidate point selection is designed to make the particles satisfy the constraints. Finally, a refinement strategy is implemented to further improve the quality of the final routing tree. Compared with other state-of-the-art algorithms, the proposed algorithm achieves a better total wirelength, which is an important index of performance, thus better satisfying the demand for delay performance of AI chip design in smart manufacturing.