Architecture Design and Performance Modeling

ML-agents for Architecture Evaluation and Design

Research Overview

We build ML agents that evaluate and guide architecture design with strong out-of-distribution generalization, clearer explanations, and higher reliability. The work combines architecture-aware mixture-of-experts models with data-driven analytical performance modeling, while improving interpretability around dependencies, bottlenecks, and critical paths. A core direction is moving from instruction-level signals to fine-grained event prediction and cycle estimation, supported by better instrumentation, scalable training pipelines, simulator integration, and LLM-based performance-agent workflows.

Research Details

Current efforts in this area include Architecture-Aware Mixture of Experts: Generalizable architecture modeling for OOD settings. , Explainable and Reliable Architecture Insight Engine: Focus on dependency/bottleneck/critical-path interpretation and trustworthy predictions. , Fine-Grained Event-to-Cycle Prediction: Move from coarse event/cycle prediction to fine-grained event pipelines for cycle determination. , and Simulator and LLM Integration: Integrate with simulators and develop LLM performance agents for architecture workflows. .