Home/News/NVIDIA HORIZON Agent Achieves 100% RTL Benchmark Completion
MarkTechPost3 min read

NVIDIA HORIZON Agent Achieves 100% RTL Benchmark Completion

NVIDIA HORIZON Agent Achieves 100% RTL Benchmark Completion

NVIDIA Research introduced HORIZON, a novel agent framework designed for hardware design that treats the process as repository-level code evolution. This approach successfully achieved 100% completion across all evaluated Register-Transfer Level (RTL) benchmark suites, according to the research team. HORIZON operates by managing hardware design problems within version-controlled repositories, rather than relying on single-turn code generation. The framework requires a structured Markdown harness as its primary input, which includes a goal, domain-knowledge directions, an evaluator specification, and an acceptance predicate. A bootstrap agent then compiles this harness into a project pack, defining the agent policy, executable evaluator, and acceptance predicate, alongside version-control policies and domain skills.

Each design problem is encapsulated within a git worktree. After the initial bootstrap, the HORIZON agent operates in a continuous loop without further human intervention. This loop involves planning a target state, editing the worktree to approach that target, invoking relevant tools, and executing an evaluator. The acceptance predicate then determines whether the changes are sufficient to commit as a new version or if the attempt has failed. For RTL design, the evaluator (E_p) can incorporate tools such as compilers, simulators, coverage extractors, and assertion checkers. The research team explicitly states that while HORIZON demonstrates significant progress, agentic hardware design is not yet a fully solved problem, acknowledging the complexities beyond plausible Verilog, including cycle-level behavior, reset conventions, and bit widths.

The structured Markdown harness, denoted as p = (π_agent, E_p, A_p, Γ_p, Ω_p), specifies the agent policy (π_agent), the executable evaluator (E_p), the acceptance predicate (A_p), the version-control policy (Γ_p), and the domain skills (Ω_p). This modular design allows HORIZON to be potentially applied to other domains beyond hardware design, where the evaluator slot could be adapted to include tools like unit tests, theorem provers, profilers, or synthesis tools. The research paper detailing HORIZON was published on arXiv.

Original source — read the full reporting at the publisher:

Read on MarkTechPost

Read next