NVIDIA HORIZON: Autonomous Agent Closes 100% of RTL Benchmarks in Chip Design

NVIDIA presented HORIZON in July 2026 — an autonomous agent for designing digital circuits at the RTL level, which was the first in the industry to achieve 100% completion of standard RTL benchmark tasks.

What is RTL and why is it difficult

RTL (Register Transfer Level — register transfer level) — a low-level description of chip logic in specialized languages: Verilog, VHDL, or SystemVerilog. At this level, engineers describe how data moves between processor or GPU registers in each cycle. Writing correct RTL code is a labor-intensive and responsible process: a single logical error becomes a hardware bug that cannot be fixed with a software update after production.

Before HORIZON, automated systems handled RTL tasks only partially: tools offered code fragments but could not independently iterate and develop the task into a fully working solution.

How HORIZON works without human intervention

The architectural foundation of the agent — each RTL task is isolated in a separate git repository through the git worktrees mechanism. The agent acts like an experienced engineer: writes code, runs verification, analyzes errors, and makes corrections — until the task is solved correctly. Operator intervention is not required at any step.

• Complete autonomy: no single point requiring human intervention
• Isolation: each task in a separate versioned repository
• Iterative verification: the agent independently fixes detected errors
• Mechanism: git worktrees for parallel and independent exploration of options
• Result: 100% completion of standard RTL benchmarks

This approach allows the agent to maintain full history of changes, revert to working states, and if necessary, simultaneously develop multiple solution hypotheses.

Why 100% on RTL benchmarks matters

One-hundred-percent completion of RTL benchmarks is a significant milestone in automating hardware design. Previously, such tasks required specialists with years of experience in digital circuit synthesis. For NVIDIA, which designs some of the world's most complex GPUs, such a tool opens the path to automating a significant portion of the development cycle for next-generation accelerators.

If results are reproduced on real commercial projects, HORIZON could shift the balance between manual RTL writing and automatic: engineers could switch from routine circuit coding to high-level architectural work.

What this means

Autonomous agents are penetrating hardware design — one of the most technically demanding engineering disciplines. NVIDIA HORIZON's result demonstrates that AI systems are already capable of closing complete RTL task cycles without human participation. The next question is how much the agent scales from academic benchmarks to real industrial projects at the level of future NVIDIA architecture generations.

Frequently asked questions

What are RTL benchmarks?

RTL benchmarks are sets of standardized tasks for describing digital circuits in Verilog or VHDL languages. They test whether an automated system can generate syntactically correct and logically sound hardware code without human intervention.

How does HORIZON differ from regular AI code assistants?

Tools like GitHub Copilot offer code fragments but do not manage the complete task solution cycle. HORIZON is an autonomous agent: it independently runs verification, interprets errors, and iterates to a working result, preserving the entire work history in isolated git worktrees.