NVIDIA Research has introduced a new framework called HORIZON that changes how hardware design works. Instead of tackling hardware as isolated tasks, HORIZON treats it like evolving code in a repository. This approach uses version control to manage hardware design problems, turning them into living projects that improve over time.
At the heart of HORIZON is the idea of hosting each design problem as a separate git worktree. This means the system manages multiple versions of designs in parallel, evolving them step-by-step. The only input it needs from users is a structured Markdown harness, which guides the process. From there, HORIZON runs a self-contained agent loop that iterates on the design automatically.
What’s impressive is how well this works in practice. In tests, HORIZON hit a 100% pass rate across every evaluated RTL benchmark suite. These benchmarks include ChipBench, RTLLM-2.0, and Verilog-Eval, plus nine verification categories from CVDP. The small misses recorded were due to specification-harness defects, not the design itself. The first iteration of the agent started with a 47.8% pass rate, showing clear progress through repeated cycles.
How HORIZON Works Under the Hood
HORIZON’s innovation lies in how it uses git histories as an experience buffer. Instead of storing past results separately, the framework treats the repository’s history as the memory of the design process. This makes each iteration aware of everything that came before it.
It also maintains a persistent model session that carries information across iterations. This continuous memory allows the agent to build on previous knowledge and avoid repeating mistakes. The backbone for these experiments is the GPT-5.3 model, which powers the reasoning and evolution within the loop.
All experiments ran on a robust AMD EPYC 9334 32-core host with 512 GB of RAM. This setup supported the heavy computational needs of evolving RTL designs and running comprehensive verifications.
Why This Matters for Hardware Design
Hardware design has traditionally required manual effort and repeated testing. HORIZON’s approach automates much of this by treating design evolution like software development. This means designers can track progress, revert changes, and explore new ideas more easily.
Despite the 100% benchmark completion, the HORIZON team is clear that agentic hardware design is not solved. The system still depends on structured inputs and can run into specification issues. But the results show a promising path forward for automating complex hardware tasks.
By framing hardware design as repository-level evolution, HORIZON changes how teams might work on chips in the future. It blends AI, version control, and hardware verification into a streamlined loop. This could speed up chip development and reduce errors.
The technical paper describing HORIZON was published on June 30th, 2026, with news coverage following on July 1st. The project was developed by NVIDIA Research, with contributions from Cunxi Yu, Chenhui Deng, Nathaniel Pinckney, and Brucek Khailany.
This new framework offers a glimpse into how AI and software tools can reshape hardware design workflows. It’s a step toward smarter, more automated chip development, driven by AI agents that learn from every iteration.
Based on
- NVIDIA HORIZON: A Hands-Free Agent that Evolves Git Worktrees and Hits 100% RTL Benchmark Completion — marktechpost.com
- A Self-Evolving Agent Framework That Treats Hardware Design as Repository-Level Code Evolution (Nvidia Research) — semiengineering.com
- NVIDIA AI Introduces ASPIRE: A Self-Improving Robotics Framework Reaching 31% Zero-Shot on LIBERO-Pro Long Tasks – mGrowTech — mgrowtech.com
- HORIZON: Agentic Hardware Design as Repository-Level Code Evolution | Nvidia Research 2026 – News and Statistics – IndexBox — indexbox.io
- Revolutionizing Hardware Design: LLMs & GNNs for HLS Optimization (2026) — sassafrasriver.org
