Modern chip design is increasingly constrained by engineering time: RTL development and verification require specialized hardware expertise, precise reasoning, and repeated interactions with EDA tools. Large language models (LLMs) have sped up code generation, and AI agents extend that benefit by using verification feedback to iteratively correct mistakes. This is especially important for RTL, where correctness depends on exact temporal behavior and many bugs only surface during tool‑based validation.
This article describes how the ACE‑RTL agent and NVIDIA Nemotron 3 Ultra work together on agentic RTL tasks. ACE‑RTL implements an iterative generate–test–reflect workflow, while Nemotron 3 Ultra adds long‑context reasoning and training aligned to RTL generation, editing, and debugging.
The CVDP benchmark
The Comprehensive Verilog Design Problems (CVDP) benchmark evaluates LLMs on realistic RTL generation, modification, debugging, and verification tasks. Unlike earlier benchmarks that used short, self‑contained Verilog prompts, CVDP better reflects practical hardware design problems. Categories include:
- cid002: RTL Code Completion
- cid003: RTL Natural Language Spec to Code
- cid004: RTL Code Modification
- cid005: RTL Spec to RTL with Module Reuse
- cid007: RTL Code Improvement (Linting/QoR)
- cid012: Testbench Stimulus Generation
- cid013: Testbench Checker Generation
- cid014: Assertion Generation
- cid016: Debugging / Bug Fixing
Why RTL requires agentic workflows
RTL coding is not merely "write Verilog from a prompt." A practical assistant must read specifications, understand existing modules, reuse and modify code, interpret simulation failures, debug mismatches, and propose targeted fixes. Engineers rarely solve complex RTL tasks in a single attempt; they iterate with compilers, simulators, lint tools, waveform inspection, and verification feedback. These naturally agentic workflows benefit from agents that can maintain context across iterations.
Inside the ACE‑RTL agent
ACE‑RTL implements a generate → run checks → analyze feedback → refine loop via three cooperating components:
- Generator: produces or updates RTL code.
- Reflector: analyzes simulation feedback, identifies likely root causes, and suggests high‑level fixes.
- Coordinator: maintains an evolving debugging context across iterations, choosing which history and feedback should guide the next generation.
This structure lets the agent build on previous failures instead of repeatedly repeating the same mistakes.
Performance on CVDP debugging (cid016)
Using ACE‑RTL noticeably improves pass rates. Examples shown include: GLM 5.2 improving from 44.0% standalone to 94.3% with ACE‑RTL; Kimi K2.6 from 68.6% to 97.1%; and Nemotron 3 Ultra from 65.7% to 100.0%. These results underscore the importance of iterative tool feedback and context evolution for realistic RTL tasks.
Fixing the agent and switching only the LLM isolates model effects: across nine CVDP categories, ACE‑RTL with Nemotron 3 Ultra achieves a 97.1% average pass rate, compared with 95.2% for Kimi K2.6 and 92.1% for GLM 5.2. Nemotron 3 Ultra reaches 100% pass rate in multiple categories, including RTL completion, spec‑to‑RTL generation, modification, module reuse, linting and QoR improvement, testbench stimulus and checker generation, assertion generation, and debugging.
Fewer tokens, greater efficiency
Nemotron 3 Ultra attains this accuracy with much lower token usage: on average it uses 6,629 tokens per iteration, versus 9,156 for GLM 5.2 and 22,579 for Kimi K2.6. That represents roughly 28% fewer tokens than GLM 5.2 and 71% fewer than Kimi K2.6 while achieving the highest average pass rate. Lower token usage reduces inference overhead for each refinement step, allowing more RTL iterations within the same compute budget and faster results for engineers.
Nemotron 3 Ultra: designed for long‑running agentic reasoning
Nemotron 3 Ultra is a 550B total‑parameter, 55B active‑parameter Mixture‑of‑Experts Hybrid Mamba‑Attention model intended for long‑running agents. It was pretrained on 20 trillion text tokens and extended to a 1M‑token context length. Among open models in its class, Nemotron 3 Ultra leads on key agentic benchmarks while maintaining a smaller active footprint.
The hybrid Mamba‑Attention reduces attention cost and KV‑cache footprint, and the MoE design improves accuracy per active parameter. According to the reported figures, this yields up to 5× higher throughput and 30% lower costs compared to other open models. For RTL workflows, where agent context grows quickly (specification, module code, generated RTL, simulator output, assertion failures, prior fixes), efficient long‑context inference is practically decisive.
Training Nemotron 3 Ultra on RTL tasks
Nemotron 3 Ultra’s RTL capability benefits from the synthetic data generation (SDG) pipeline described in the ACE‑RTL work. The pipeline aims to produce high‑quality RTL training data that reflects how hardware engineers and RTL agents actually work: generating code from specs, editing existing implementations, and debugging failing designs.
The SDG pipeline starts from high‑quality RTL seed designs filtered from public repositories. Open LLMs convert these seeds into instruction‑style training samples, guided by a few in‑context examples that define the expected format and task style. For spec‑to‑RTL tasks, LLMs synthesize natural‑language specifications paired with golden RTL implementations derived from the seeds.
A key innovation is task diversity: beyond spec‑to‑RTL, the pipeline creates RTL code editing and debugging samples. For editing, the original seed is the golden implementation while a simplified version becomes the input; the paired spec describes the feature extension or corner case needed to recover the golden RTL. For debugging, realistic RTL errors are injected (incorrect FSM transitions, handshake or timing violations, other common mistakes), and the corresponding specification includes diagnostic information like observed failure behavior or signal mismatches so the model learns to repair faulty RTL using code context and feedback.
All generated samples pass through a data filter before training: syntax checks, benchmark decontamination, and rubric‑based scoring using an LLM‑as‑judge to assess semantic alignment between specification and implementation. This filtering helps ensure the synthetic data is structurally valid and useful for training.
Training on editing and debugging tasks exposes Nemotron 3 Ultra to the same types of repair and refinement problems it faces inside an agentic RTL workflow, which is critical for iterative refinement scenarios.
Getting started and integrations
Nemotron 3 Ultra is an open model available today and can be integrated into engineering toolchains. The announcement highlights partnerships and integrations such as:
- Cadence: where ChipStack AI SuperAgent is paired with Nemotron 3 Ultra for front‑end design and verification agent orchestration; other SuperAgents (InnoStack, ViraStack, AuraStack) are also mentioned.
- Siemens: Nemotron 3 Ultra is used with the Questa One Agentic Toolkit and integrated into Fuse EDA AI Agent to provide up to 5×–10× token efficiency improvements.
- Synopsys AgentEngineer: leverages Nemotron 3 Ultra models in a multi‑agent system to pursue autonomous verification closure and continuous feedback loops, reducing manual verification effort.
Conclusion
Combined with the ACE‑RTL agentic workflow and an RTL‑oriented synthetic training pipeline, Nemotron 3 Ultra delivers high pass rates across realistic RTL tasks while using substantially fewer tokens per iteration than comparable models. The model architecture, long‑context support, and task‑diverse training data make Nemotron 3 Ultra particularly suited to agentic RTL development workflows.
