Presentation
CANCELED - Early IR Drop Prediction using Machine Learning for Power Grid
DescriptionMeeting voltage drop specification has been an increasing challenge with process node shrinking, increase in layer stack and increase in power consumption. In custom design, power grid generation are not IR driven today. Grid gets implemented in layout first and then verified by tools on its robustness. This involves multiple iterations sometimes in advanced stage of design milestones which creates challenges in implementing layout edits as well as disturbs converged parameters and impact schedules.
We propose a machine learning (ML) solution which works at the power grid definition phase of a custom IP, it tries to predict block level IRdrop on a given specification when no layout and placement information is available. It uses physical and electrical specifications as features to train the ML model. User can tune metal layer, width & pitch values to derive optimal power grid to proceed for physical implementation. The training data sets are design testcases run with IR analysis tools. Appropriate features like block dimension, metal layers, width, pitch of power grid are extracted along with IRdrop results for the testcases.
We propose a machine learning (ML) solution which works at the power grid definition phase of a custom IP, it tries to predict block level IRdrop on a given specification when no layout and placement information is available. It uses physical and electrical specifications as features to train the ML model. User can tune metal layer, width & pitch values to derive optimal power grid to proceed for physical implementation. The training data sets are design testcases run with IR analysis tools. Appropriate features like block dimension, metal layers, width, pitch of power grid are extracted along with IRdrop results for the testcases.
Event Type
Engineering Track Poster
TimeTuesday, July 11th5:00pm - 6:00pm PDT
LocationLevel 2 Exhibit Hall
Back-End Design
Embedded Systems
Front-End Design
IP
RISC-V