| Summary: | Due to recent aggressive process scaling into the nanometer regime, power delivery network design faces many challenges that set more stringent and specific requirements to the EDA tools. For example, from the perspective of analysis, simulation efficiency for large grids must be improved and the entire network with off-chip models and nonlinear devices should be able to be analyzed. Gated power delivery networks have multiple on/off operating conditions that need to be fully verified against the design requirements. Good power delivery network designs not only have to save the wiring resources for signal routing, but also need to have the optimal parameters assigned to various system components such as decaps, voltage regulators and converters. This dissertation presents new methodologies to address these challenging problems.
At first, a novel parallel partitioning-based approach which provides a flexible network partitioning scheme using locality is proposed for power grid static analysis. In addition, a fast CPU-GPU combined analysis engine that adopts a boundary-relaxation method to encompass several simulation strategies is developed to simulate power delivery networks with off-chip models and active circuits. These two proposed analysis approaches can achieve scalable simulation runtime.
Then, for gated power delivery networks, the challenge brought by the large verification space is addressed by developing a strategy that efficiently identifies a number of candidates for the worst-case operating condition. The computation complexity is reduced from O(2^N) to O(N).
At last, motivated by a proposed two-level hierarchical optimization, this dissertation presents a novel locality-driven partitioning scheme to facilitate divide-and-conquer-based scalable wire sizing for large power delivery networks. Simultaneous sizing of multiple partitions is allowed which leads to substantial runtime improvement. Moreover, the electric interactions between active regulators/converters and passive networks and their influences on key system design specifications are analyzed comprehensively. With the derived design insights, the system-level co-design of a complete power delivery network is facilitated by an automatic optimization flow. Results show significant performance enhancement brought by the co-design.
|
|---|
