CHICO-Agent: A Cross-Layer Optimization Framework for Chiplet Systems Based on Large Language Models

This article introduces CHICO-Agent—a cross-layer optimization framework for Chiplet systems based on Large Language Models (LLMs). Through a persistent knowledge base and multi-agent workflow, this framework addresses the complexity of cross-layer collaborative optimization across application, architecture, chip, and packaging layers in 2.5D/3D Chiplet system design, and can find lower-cost configurations compared to traditional baseline methods like simulated annealing.

The growth of large language models and AI workloads has pushed traditional monolithic silicon chips to their physical and economic limits. Chiplet technology, which splits large chips into smaller ones and reintegrates them, offers advantages such as improved yield, cost optimization, and enhanced flexibility, but it also drastically increases the complexity of cross-layer optimization: decisions across application layer (workload characteristics), architecture layer (computing unit ratio, etc.), chip layer (process selection, etc.), and packaging layer (2.5D/3D selection, etc.) need to be coordinated, forming a combinatorial explosion of design space that traditional methods struggle to handle.

CHICO-Agent adopts a multi-agent collaboration architecture:

• Knowledge Base: Persistently records historical design decisions, parameter-performance correlation patterns, and design rules, accumulating experience as optimization progresses.
• Admin Agent: Central coordinator that decomposes goals, assigns tasks, integrates feedback, and updates the knowledge base.
• Domain Agents: Specialized agents for each layer of application, architecture, chip, and packaging.

The optimization workflow consists of five phases: Requirement Analysis (application agent extracts workload requirements) → Architecture Exploration (generates candidate architectures) → Chiplet Decomposition (maps architecture to Chiplets) → Packaging Optimization (plans physical implementation) → Evaluation and Iteration (simulation evaluation and coordination for the next round of optimization).

CHICO-Agent's uniqueness lies in:

1. Parameter-Result Correlation Learning: Achieves cross-iteration experience accumulation through the knowledge base, recording effective parameter combinations and correlation patterns, avoiding the "stateless" start-from-scratch approach of traditional methods.
2. Natural Language Reasoning: Uses LLM capabilities to understand natural language constraints, analyze the pros and cons of solutions, and diagnose anomalies.
3. Interpretable Design Trajectory: Each decision has a natural language explanation, allowing traceability of reasons and facilitating manual review.

The research team validated the effectiveness of CHICO-Agent in an AI accelerator scenario:

• Comparison Baselines: Simulated Annealing (SA), Bayesian Optimization (BO), Genetic Algorithm (GA).
• Results: 18% lower cost than SA, 12% lower than BO, and 15% lower than GA; faster convergence speed (50-100 rounds vs SA's 200+ rounds); design quality more aligned with engineering practices.
• Case Study: In Transformer accelerator design, CHICO-Agent avoided SA's local optimal traps (e.g., interconnection bandwidth bottlenecks, power overrun).

CHICO-Agent represents trends in the EDA field:

• LLMs as Design Partners: Understand high-level intent, perform common-sense reasoning, learn from experience, and provide interpretable suggestions.
• Cross-Layer Collaboration: Multi-agent architecture naturally decomposes complex cross-layer problems.
• New Human-Machine Collaboration Model: Enhances designers' capabilities, allowing them to focus on high-level decisions while leaving tedious tuning to AI.

Limitations of CHICO-Agent: Dependence on external simulation (high cost), unproven knowledge generalization, and insufficient real-time performance. Future directions:

• Train surrogate models to replace expensive simulations for accelerated evaluation;
• Online learning of actual chip performance after deployment;
• Support multi-objective optimization (Pareto front exploration);
• Automatically generate test cases to verify design correctness.