Industry Track: Trajectory.AI: An Automated P&R Flow Trajectory Agent

A synthesis and place-and-route (P&R) tool is a complex engineering workflow that involves hundreds of steps, each doing an atomic task sequentially to produce a final netlist. Oftentimes, a flow step that produces suboptimal quality of results (QoRs) can impact steps downstream. A time series of QoR data produced at each step of the design flow, referred to as a "trajectory," has been used traditionally for understanding and debugging the flow anomalies. However, the process involves manual inspection of trajectory anomalies through utility scripts, and recommendations are done based on the review and analysis of experienced engineers.We propose Trajectory Agent, which is built into the tool to collect, analyze, and detect trajectory-related suboptimalities with an explainable framework. Trajectory Agent automates the process of finding trajectory anomalies, and root causing to a step that underperforms in the context of multiple tradeoff objectives like total negative slack, area, and power. This is achieved by learning essential patterns and statistics from a large number of trajectories, accumulated over many runs from different designs and flow variations over months.Specifically, we characterize a part of the design flow by modeling the QoR delta between start and end steps with a multivariate Gaussian mixture model. The modeling can be performed on any flow segments of interest. The multivariate model captures the interactions of different QoR metrics (e.g., power, area, timing, and congestion), which are typically correlated due to the tradeoffs in various optimization steps. The mixture of Gaussian models accounts for different design characteristics and different flow variations due to constraints and flow configurations, while preserving simplicity and explainability. This model embodies a full probabilistic description of the flow segment, and hence can be used for QoR prediction, anomaly detection and root causing.Trajectory Agent makes it possible to observe and analyze the deviation of a QoR trajectory early in the flow due to possible step ineffectiveness and/or specific design characteristics, and take actions accordingly that adjust the flow on the fly. It is the foundation of building a flow recommendation system that dynamically caters to different designs and flow variations based on the time series of trajectory data.