AI Evolves Better Computer Chip Designs

A new artificial intelligence system can automatically generate and improve computer chip designs, achieving near-perfect accuracy while optimizing for power consumption and performance. This breakthrough addresses a critical bottleneck in electronics manufacturing, where manual design processes are increasingly time-consuming and error-prone.

Researchers developed REvolution, a framework that combines large language models with evolutionary computation to automatically create Register-Transfer Level (RTL) code—the blueprint language used to design computer chips. The system evolves a population of design candidates through multiple generations, systematically improving both functional correctness and hardware efficiency metrics.

The method works by maintaining two separate populations of design candidates: one for designs that fail functional testing and another for designs that work correctly but need optimization. For failed designs, the system focuses on fixing errors, while for working designs, it concentrates on improving power consumption, performance, and physical area. The framework uses six different prompt strategies to guide the evolutionary process, including fixing errors, simplifying code, exploring new approaches, refactoring existing designs, improving performance, and combining successful elements from multiple designs.

Experimental results demonstrate significant improvements. On standard hardware design benchmarks, REvolution increased the pass rate—the percentage of designs that function correctly—by up to 24 percentage points across different AI models. The system achieved a 95.5% pass rate using DeepSeek-V3, comparable to state-of-the-art methods but without requiring specialized training or external tools. The evolved designs also showed substantial hardware improvements, with power consumption reductions up to 67% and clock period improvements up to 48% compared to initial designs.

This advancement matters because modern electronics rely on increasingly complex integrated circuits, and manual RTL design has become a major bottleneck prone to human error, high costs, and lengthy development times. Current AI approaches for chip design often get stuck with suboptimal solutions because they focus on refining a few candidates rather than exploring the full design space. REvolution's evolutionary approach enables broader exploration, moving beyond local optima to discover globally better designs.

The system does have limitations. Its performance depends on the underlying AI model's capabilities, as demonstrated by the performance gap between different models. Additionally, the framework requires multiple generations of evolution, which involves computational costs for simulation and synthesis. The current implementation also focuses on digital circuit design and may need adaptation for other types of hardware design.

REvolution represents a shift from previous methods that relied on local search and iterative refinement. By combining AI's generative capabilities with evolutionary computation's systematic exploration, the framework can discover highly optimized hardware designs that might otherwise remain undiscovered. This approach could accelerate electronics development while maintaining or improving design quality, potentially impacting everything from consumer devices to specialized computing hardware.