A new open-source framework called CHIA is making it possible for AI agents to design computer chips and their software in tandem, automating one of the most complex and labor-intensive tasks in technology. The framework, introduced by researchers working across computer architecture and systems design, treats the construction and deployment of AI-driven design workflows as a core objective, enabling researchers to build sophisticated hardware and software development processes that previously required extensive manual engineering.

What Problem Does CHIA Actually Solve?

Designing modern computer systems is extraordinarily difficult. Architects, system designers, compiler engineers, and specialists in chip manufacturing must navigate enormous design spaces while reasoning across multiple layers of abstraction, from software down to the physical implementation of circuits. Until now, applying AI to this work has been limited to isolated experiments on small-scale problems, largely because there was no good way to express and reliably deploy complex AI-driven design workflows at scale.

The traditional approach of writing custom scripts to glue together different design tools breaks easily and doesn't scale well when you're working across hundreds of different machines with varying hardware configurations. Alternatively, letting AI agents run completely unsupervised inside design frameworks creates verification and validation nightmares, making it difficult to ensure the designs actually work correctly.

CHIA bridges this gap by providing a structured way to express AI-driven design flows while maintaining the rigorous testing and validation that hardware design demands. The framework is built on the premise that the design flow itself should be treated as a first-class objective, not an afterthought.

How Does CHIA Enable AI Agents to Design Hardware?

At its core, CHIA works through what researchers call "CHIA loops," which are directed cyclic graphs whose nodes execute various design tools, simulators, AI models, and evolutionary coding agents. Think of it as a sophisticated assembly line where different specialized tools and AI systems work together in a coordinated sequence, with feedback loops that allow the system to iterate and improve.

The CHIA library includes pre-built node implementations for many widely used tools in chip and system design, including Chipyard, gem5, ChampSim, FireSim, Hammer, Vivado, AlphaEvolve, and AdaEvolve. This means researchers can compose sophisticated design flows from reusable components rather than building everything from scratch.

Beyond just connecting tools together, CHIA provides essential features for conducting rigorous research around these workflows:

• Isolation: AI models are kept separate from hardware tools to prevent unexpected interactions and ensure reproducibility.
• Profiling and Monitoring: Detailed tracking mechanisms allow researchers to understand how the design process is performing and where bottlenecks occur.
• Fault Tolerance: The system can handle failures gracefully and continue operating reliably across hundreds of heterogeneous systems, including CPUs, FPGAs, and GPUs, whether running on public cloud infrastructure or on-premises hardware.

What Can AI Agents Actually Accomplish With CHIA?

The researchers demonstrated CHIA's capabilities through five concrete case studies, each showing different applications of agentic AI to hardware and software design. These weren't theoretical exercises; they produced real, working designs that meet the exacting standards of the hardware industry.

One particularly impressive achievement involved AI agents implementing microarchitectural features in hardware description language (RTL). The designs created by these agents delivered substantial performance improvements while successfully executing all 25 trillion or more instructions from the SPEC CPU2006 reference suite, a standard benchmark used to validate processor designs. These agent-written implementations improved performance while meeting or improving frequency and area constraints in both open-source and commercial chip design kits.

The five case studies included automatic alignment between RTL and simulator specifications, LLM-driven implementation of microarchitectural features, optimization of critical paths in processor design with awareness of instructions per cycle (IPC), evolutionary discovery of new processor architectures, and even agentic fixing of issues reported on GitHub by maintainers.

Why Does This Matter for the Future of Chip Design?

The implications are significant. Recent advances in agentic artificial intelligence have already demonstrated promising results in hardware design, from automatically generating state-of-the-art cache replacement policies to generating complete RISC-V microprocessor designs from scratch. CHIA provides the infrastructure to scale these capabilities and enable more researchers to explore how AI can accelerate innovation in computer architecture and systems design.

The framework acknowledges a fundamental reality: humans, AI agents, and existing hardware design tools must all work together smoothly. CHIA makes that collaboration possible by providing a principled, scalable way to orchestrate complex workflows without sacrificing the verification and validation rigor that hardware design requires. By open-sourcing the framework, the researchers hope to enable the broader community to build on these foundations and explore new applications of agentic AI to hardware and software co-design challenges.