chia-agentic-hardware-software-codesign

name	chia-agentic-hardware-software-codesign
description	CHIA framework for principled agentic AI-driven hardware/software co-design. Treat hardware/software design flows as directed cyclic graphs (CHIA loops) with nodes executing SoC design tools, simulators, AI models, and evolutionary agents. Supports Chipyard, gem5, ChampSim, FireSim, Vivado, AlphaEvolve, and more. Enables isolation, profiling, fault-tolerant execution, and reliability across heterogeneous systems.
activation_keywords	["CHIA","agentic AI","hardware software co-design","computer architecture","systems engineering","RTL generation","gem5 simulation","Chipyard","FireSim","Vivado","AlphaEvolve","evolutionary agents","microarchitectural optimization","IPC optimization","critical path optimization","GitHub issue automation","硬件软件协同设计","智能体驱动设计","系统工程","计算机架构"]
arxiv_id	2606.2735
authors	Angela Cui, Ferran Hermida-Rivera, Jack Toubes, Raghav Gupta, Jim Fang, Chengyi Lux Zhang, Ella Schwarz, Junha Kim, Yakun Sophia Shao, Borivoje Nikolic, Christopher W. Fletcher, Sagar Karandikar
published	2026-06-25T00:00:00.000Z
categories	["cs.AR","cs.AI","cs.SE"]

CHIA: Agentic AI-Driven Hardware/Software Co-Design Framework

Core Contribution

CHIA (Agentic AI Hardware/Software Co-design) is an open-source framework that enables principled research on AI-driven hardware/software co-design. It treats the construction and deployment of co-design flows as a first-class objective.

Key Innovation: CHIA Loops

CHIA expresses agentic AI-driven hardware and software design flows as directed cyclic graphs called CHIA loops. Each node executes:

• System-on-chip design tools (Chipyard, Hammer, Vivado)
• Microarchitectural simulators (gem5, ChampSim)
• Software build systems
• AI models (LLMs, evolutionary agents)
• Evolutionary coding agents (AlphaEvolve, AdaEvolve)

Framework Components

1. CHIA Library

Provides node implementations for popular tools:

• Chipyard: Agile SoC design framework
• gem5: Architectural simulation
• ChampSim: Cache simulator
• FireSim: FPGA-accelerated simulation
• Hammer: Physical design (commercial CAD tools)
• Vivado: FPGA synthesis
• AlphaEvolve: LLM-based evolutionary agents
• AdaEvolve: Adaptive evolutionary agents

2. Principled Science Features

• Isolation: Separates AI models from hardware tools
• Profiling: Performance monitoring mechanisms
• Fault tolerance: Robust execution handling
• Scalability: Reliability across hundreds of heterogeneous systems
• Multi-platform: CPUs, FPGAs, GPUs across cloud/on-prem

Case Studies

1. RTL-to-gem5 Simulator Alignment

Automatic generation and alignment of RTL with gem5 simulators for consistency verification.

2. LLM-Driven Microarchitectural Feature Implementation

Agentic AI automatically implements microarchitectural features in RTL code.

3. IPC-Aware Critical Path Optimization

Intelligent optimization of instruction-level parallelism aware of critical paths.

4. Evolutionary Architectural Discovery

Evolutionary agents discover novel architectural configurations.

5. Maintainer-Friendly GitHub Issue Fixing

Agentic automation for resolving hardware/software design issues.