ACE-RTL

ACE-RTL is an agentic system for RTL code generation, verification, and iterative repair. It coordinates three core components, a Generator, a Reflector, and a Coordinator, through an agentic context evolution workflow that continuously refines the design based on tool feedback and prior attempts.

The release includes the agent skills, reusable role components, benchmark integration scripts, and support for external LLM backends. CVDP is included as a representative benchmark to demonstrate the effectiveness of ACE-RTL, rather than as the sole target of the framework. By default, the external LLM path uses the publicly available NVIDIA Inference API through an NVIDIA-supported Python client interface.

The default hosted model is nvidia/nemotron-3-ultra-550b-a55b through https://integrate.api.nvidia.com/v1. API keys are read only from environment variables and must never be committed to this repository. Users can also set ACE_RTL_LLM_SCRIPT to route calls through their own adapter.

Requirements

• Python 3.11 recommended
• Python packages listed in the committed requirements.txt
• git, make, curl
• CVDP simulation backends required by the selected rows, commonly iverilog, vvp, verilator, and yosys
• Commercial EDA tools only for CVDP rows that explicitly require them, for example Cadence Xcelium/IMC coverage flows
• A NVIDIA Inference API key in NVIDIA_API_KEY

The top-level requirements.txt is the authoritative Python dependency manifest for ACE-RTL setup and dependency scanning.

Setup

Run from this repository root:

./skills/ace-rtl/scripts/setup_cvdp_native_env.sh

The setup helper creates .cvdp-native if needed, clones https://github.com/NVlabs/cvdp_benchmark into external/cvdp_benchmark if it is not already present, installs the Python packages from the top-level requirements.txt, and downloads CVDP JSONL files from Hugging Face repo nvidia/cvdp-benchmark-dataset. Existing datasets are reused.

First check whether native simulator tools are already available in the shell that will run ACE-RTL:

command -v iverilog vvp verilator yosys make

If your site provides tools through modules, Conda/Mamba, or installation roots such as IVERILOG_HOME, load that environment first and rerun the check. Install native simulator tools only if they are still unavailable. On Ubuntu/Debian this is typically:

sudo apt-get update
sudo apt-get install -y iverilog verilator yosys make

For commercial/Cadence CVDP rows, configure your site-provided EDA tools before running the evaluator. Start from skills/eda_tool_setup/SKILL.md and the example script at skills/eda_tool_setup/scripts/setup_eda_env.example.sh.

Set your NVIDIA Inference API key in the shell that will run ACE-RTL:

export NVIDIA_API_KEY="<your NVIDIA Inference API key>"

Optional LLM defaults:

export ACE_RTL_MODEL="nvidia/nemotron-3-ultra-550b-a55b"
export ACE_RTL_MAX_TOKENS=32768
export ACE_RTL_ENABLE_REASONING=1
export ACE_RTL_REASONING_BUDGET=32768
export ACE_RTL_STREAM=1
export ACE_RTL_TEMPERATURE=0.2
export ACE_RTL_TOP_P=0.95

To use a non-default provider, implement a small adapter script and set:

export ACE_RTL_LLM_SCRIPT="/path/to/your_llm_adapter.py"

The adapter receives a JSON request on stdin with model, messages, temperature, top_p, max_tokens, stream, and extra_body. It may return either raw assistant text or JSON containing {"content": "..."} or an chat-completions-style choices[0].message.content payload.

Run With A Coding Agent Recommended

Use a coding agent from this repository root and ask it to use the ACE-RTL skill. Example prompt:

Use ACE-RTL to run CVDP cid016 from
datasets/cvdp-benchmark-dataset/cvdp_v1.1.0_nonagentic_code_generation_no_commercial.jsonl
with generator=nvidia/nemotron-3-ultra-550b-a55b,
reflector=nvidia/nemotron-3-ultra-550b-a55b,
coordinator=nvidia/nemotron-3-ultra-550b-a55b.
Use external/cvdp_benchmark as the official CVDP checkout.

The agent will read skills/ace-rtl/SKILL.md, inspect the selected CVDP rows, verify local harness tools, choose the native or Docker path, and preserve run artifacts under outputs/.

Command To Run

For a direct non-agentic, no-commercial CVDP run, set --cid to the target category:

export NVIDIA_API_KEY="<your NVIDIA Inference API key>"
export CVDP_NATIVE_HARNESS=1

# The native runner uses open-source simulator tools from this shell.
# Load your site/module/Conda environment first, or prepend a tool root.
# The run will preflight required tools such as iverilog/vvp before evaluation.
command -v iverilog vvp

.cvdp-native/bin/python3 skills/ace-rtl/scripts/ace_cvdp_native_runner.py \
  --workspace "$PWD" \
  --cvdp-repo "$PWD/external/cvdp_benchmark" \
  --dataset "$PWD/datasets/cvdp-benchmark-dataset/cvdp_v1.1.0_nonagentic_code_generation_no_commercial.jsonl" \
  --cid cid016 \
  --generator-model nvidia/nemotron-3-ultra-550b-a55b \
  --reflector-model nvidia/nemotron-3-ultra-550b-a55b \
  --coordinator-model nvidia/nemotron-3-ultra-550b-a55b \
  --attempts 5 \
  --datapoint-workers 5 \
  --max-iterations 30

If command -v iverilog vvp fails but the tools are installed under explicit roots, set those roots before launching the runner:

export IVERILOG_HOME=/path/to/iverilog
export VERILATOR_HOME=/path/to/verilator
export YOSYS_HOME=/path/to/yosys
export PATH="$IVERILOG_HOME/bin:$VERILATOR_HOME/bin:$YOSYS_HOME/bin:$PATH"
command -v iverilog vvp verilator yosys

ACE-RTL does not install simulator binaries during a run. It uses tools already discoverable from the environment and writes native_tool_preflight.json in the run directory showing which tools were required, found, or missing.

Use --attempts 1 --max-iterations 1 --limit 1 first for a shakedown if you are setting up a new machine.

Results are written under outputs/, including run.log, summary.json, and results.md.

Citation

Please cite our paper if you find ACE-RTL useful:

@article{deng2026ace,
  title={ACE-RTL: When Agentic Context Evolution Meets RTL-Specialized LLMs},
  author={Deng, Chenhui and Yu, Zhongzhi and Liu, Guan-Ting and Pinckney, Nathaniel and Khailany, Brucek and Ren, Haoxing},
  journal={NSF Workshop on Agents for Chip Design Automation (Agent4Chip)},
  year={2026}
}