Presentation
LSTM-Based Performance Modeling for Enhanced Peak Memory Prediction in Layout Versus Schematic Verification Tools
DescriptionPerformance modeling of software tools is crucial for optimizing resources and costs, yet precise predictions are challenging due to vast configuration spaces and limited data. While machine learning offers solutions, it demands extensive data collection. This paper presents a novel methodology using Long Short-Term Memory (LSTM) models to predict peak memory in VLSI layout verification tools by analyzing memory consumption patterns throughout runtime. Testing on Calibre's nmLVS tool achieved 72% average prediction accuracy on unseen data, reaching 98% in specific cases, representing a 10% improvement over traditional neural networks. This approach demonstrates that leveraging sequential performance data enhances software behavior understanding while reducing the need for large, diverse datasets.
Event Type
Work in Progress
TimeMonday, July 276:56pm - 6:57pm PDT
LocationExhibit Hall