Physics-Based Electromigration Modeling and Analysis and Optimization
- Author(s): Sun, Zeyu
- Advisor(s): Tan, Sheldon X.-D.
- et al.
Long-term reliability is a major concern in modern VLSI design. Literature has shown that reliability gets
worse as technology advances. It is expected that the future VLSI systems would have shorter reliability-induced lifetime comparing with previous generations. Being one of the most serious reliability effects, electromigration (EM) is a physical phenomenon of the migration of metal atoms due to the momentum exchange between atoms and the conducting electrons. It can cause wire resistance change or open circuit and result in functional failure of the circuit. Power-ground networks are the most
vulnerable part to EM effect among all the interconnect wires since the
current flow on this part is the largest on the chip. With new generation of
technology node and aggressive design strategies, more accurate and
efficient EM models are required. However, traditional EM approaches are
very conservative and cannot meet current aggressive design strategies.
Besides circuit level, EM also need to be
thoroughly studied in system level due to limited power and
temperature budgets among cores on chip.
This research focuses on developing physical level EM model for VLSI circuits and system level EM optimization for multi-core systems in order to overcome the aforementioned problems. Specifically, for physical level, we develop two EM immortality check methods and a power grid EM check method. Firstly, a voltage based EM immortality analysis has been developed. Immortality condition in nucleation phase can be determined fast and accurately for multi-segment interconnect wires. Secondly, a saturation volume based incubation phase immortality check method has been proposed. This method can further reduce the redundancy in VLSI circuit design by immortality check in multiphase. Furthermore, both immortality check methods are integrated into a new power grid EM check methodology (EMspice) as filter for EM analysis. These filters can accelerate the simulation by filtering out immortal trees so that we only need to do simulation on fewer trees that are mortal. Coupled EM simulation considering both hydrostatic stress and electronic current/voltage in the power grid network will be applied to these mortal trees. This tool can work seamlessly with commercial synthesis flow.
Besides physical level reliability models, system level reliability optimization is also discussed in this research. A deep reinforcement learning based EM optimization has been proposed for multi-core system. Both long term reliability effect (hard error) and transient soft error are considered. Energy can be optimized with all the reliability and other constraints fast and accurately compared to existing reliability management techniques. Last but not least, a scheduling based reliability optimization method for multi-core systems has been proposed. NBTI, HCI and EM are considered jointly. Lifetime of the system can be improved significantly compared to traditional methods which mainly focus on utilization.