NVIDIA Research Unveils HORIZON: A Groundbreaking AI Agent Achieving 100% RTL Benchmark Completion

NVIDIA Research has unveiled HORIZON, a groundbreaking hands-free agent framework that achieved a remarkable 100% completion rate across all evaluated Register-Transfer Level (RTL) benchmark suites. This significant milestone for autonomous hardware design agents signals a new era in electronic design automation (EDA), where AI takes on increasingly complex tasks.

HORIZON distinguishes itself by treating each RTL design problem as repository-level code evolution, hosted within an isolated Git worktree. The agent operates autonomously, guided by a structured Markdown harness that compiles into a project pack defining the goal, domain knowledge, an executable evaluator, and an acceptance predicate. This innovative approach allows the agent to iteratively evolve the design, committing new versions only when an acceptance gate passes, effectively using Git as its primary interface and an experience buffer for learning. The agent successfully navigated benchmarks including ChipBench, RTLLM-2.0, and Verilog-Eval, with the sole reported "miss" attributed to a benchmark defect rather than agent failure.

This achievement dramatically advances the vision of autonomous chip design. While previous multi-AI agents have shown promising results, such as 94.2% functional correctness on some benchmarks, HORIZON's perfect completion rate on a broad set of RTL challenges demonstrates a significant leap in reliability and capability. It underscores NVIDIA's broader strategy of integrating AI into every stage of chip design, from layout optimization to verification, to accelerate development cycles and achieve better-than-human results. However, the research team prudently emphasizes that agentic hardware design is "not solved," acknowledging remaining hurdles like reward hacking and long-turnaround rewards. Nevertheless, HORIZON's success portends a future where AI agents become indispensable co-pilots, freeing human engineers to focus on higher-level architectural innovation and complex problem-solving, rather than the meticulous, iterative tasks that have long characterized RTL development. The ongoing evolution of such AI frameworks promises to reshape the semiconductor industry, enabling faster, more efficient, and potentially more innovative chip designs.

Watch (Shorts)