Machine Learning-enhanced Multi-dimensional Co-Optimization of Sub-10nm Technology Node Options

This paper presents a next generation multi-objective co-optimization approach that enables the joint utilization of machine and human intelligence to address the growing challenges in all stages of technology enablement. This approach is utilized in exploring optimal combinations of a rich set of design, technology and ingredients (DTI) to push the power/performance/area (PPA) envelope across sub-10nm derivatives of Intel technology nodes and various intellectual properties (IPs). It focuses on assisting design engineers in exploring the multi-faceted design space of ever-increasing complexity that is governed by convoluted trade-offs, which directly determine end product quality. This approach is applicable to all stages of the technology definition and readiness cycle. Here, a 10× improvement in turn-around time (TAT) with better quality of results (QoR) is reported compared to purely human-optimized high-performance CPU implementations in Intel’s 10nm technology node using various transistor, standard cell, metal stack, track pattern and methodology options.