NVIDIA's Nemotron 3 Ultra model achieves top accuracy and efficiency in agentic RTL coding, surpassing other open models. The model leverages LLMs and AI agents to iteratively correct errors using verification feedback, addressing engineering time constraints in chip design.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
NVIDIA's Nemotron 3 Ultra model achieves a 97.1% average pass rate in agentic RTL coding tasks, surpassing GLM 5.2 and Kimi K2.6, while reducing token consumption per iteration by 28%-71%, demonstrating a combination of high precision and efficiency.
Modern chip design is constrained by engineering time; RTL development and verification require specialized hardware knowledge, precise reasoning, and iterative interaction with EDA tools. LLMs accelerate code generation, but AI agents further extend their impact by leveraging verification feedback for iterative error correction. For RTL, correctness depends on precise timing behavior, and many errors are only exposed during tool-based verification.
The ACE-RTL agent employs a generate-test-reflect iterative workflow: a generator produces or updates RTL code, a reflector analyzes simulation feedback and identifies root causes, and a coordinator maintains an evolving debug context across iterations. This design enables the agent to learn from previous failures rather than repeating similar errors.
In the CVDP benchmark covering nine RTL task categories, the ACE-RTL agent with Nemotron 3 Ultra achieves a 97.1% average pass rate, higher than Kimi K2.6's 95.2% and GLM 5.2's 92.1%. Nemotron 3 Ultra reaches 100% pass rates in multiple categories.
More importantly, Nemotron 3 Ultra achieves the highest accuracy while consuming significantly fewer tokens: an average of 6,629 tokens per iteration, 28% fewer than GLM 5.2's 9,156 and 71% fewer than Kimi K2.6's 22,579. This combination of high accuracy and low token cost allows more RTL tasks to be attempted within the same compute budget, and engineers can obtain results faster.
Nemotron 3 Ultra is a hybrid Mamba-Attention MoE model with 550B total parameters and 55B active parameters, designed for long-running agents. It is pre-trained on 20 trillion text tokens and supports a 1M token context length. The hybrid Mamba-Attention architecture improves throughput efficiency by reducing attention costs and KV cache footprint, while the MoE design enhances accuracy per active parameter, enabling up to 5x throughput and 30% cost reduction compared to other open models.
For RTL workflows, efficiency is critical as agent contexts grow rapidly. A single debug iteration may include specification, module code, generated RTL, simulation output, assertion failures, and previous fix attempts. Efficient context reasoning directly impacts the practicality of these agents in real engineering.
Nemotron 3 Ultra's RTL capabilities benefit from the synthetic data generation pipeline proposed in the ACE-RTL paper. This pipeline aims to create high-quality RTL training data that reflects how hardware engineers and RTL agents actually work: not only generating code from specifications but also modifying existing implementations and debugging failing designs.
This article's information comes from NVIDIA's official technical blog, a first-party source. All performance data are NVIDIA's claims under specific benchmarks (CVDP) and agent framework (ACE-RTL) and have not been independently verified by third parties.
Nemotron 3 Ultra achieves dual leadership in accuracy and efficiency in agentic RTL coding, supported by its hybrid Mamba-Attention MoE architecture and synthetic data training strategy.
Primary report
Primary source