NVIDIA Nemotron 3 Ultra Leads Open Models on Accuracy and Efficiency in Agentic RTL Coding NVIDIA Nemotron 3 Ultra 在代理式 RTL 编码的准确性和效率方面领先开源模型

TL;DR

• The ACE-RTL agent combined with NVIDIA Nemotron 3 Ultra achieves state-of-the-art accuracy (97.1% average pass rate) and efficiency (up to 71% fewer tokens per iteration) on agentic RTL coding tasks compared to models like GLM 5.2 and Kimi K2.6.
• The hybrid Mamba-Attention Mixture-of-Experts architecture of Nemotron 3 Ultra, trained on a rubric-filtered synthetic RTL dataset, enables high-throughput, low-cost, long-context inference suitable for integration with EDA tools.
• The iterative generate-test-reflect workflow of the ACE-RTL agent significantly improves pass rates across all evaluated models by leveraging tool feedback and evolving context, with Nemotron 3 Ultra reaching 100% pass rate on several CVDP categories.

Why It Matters

This development is highly relevant to AI practitioners and the hardware industry as it demonstrates a practical path to accelerating RTL design and verification—a critical bottleneck in modern chip development. By combining specialized model architectures with agentic workflows, it reduces engineering time and cost while improving correctness, setting a new benchmark for AI-assisted hardware design.

Technical Details

• The Comprehensive Verilog Design Problems (CVDP) benchmark evaluates nine agentic RTL task categories including code completion, spec-to-code generation, modification, debugging, and testbench generation, reflecting real-world hardware design workflows.
• ACE-RTL agent architecture consists of three components: a generator (produces/updates RTL), a reflector (analyzes simulation feedback and identifies root causes), and a coordinator (maintains evolving debugging context across iterations).
• Nemotron 3 Ultra uses a 550B total-parameter, 55B active-parameter Mixture-of-Experts Hybrid Mamba-Attention model with 1M-token context length, pretrained on 20 trillion text tokens and fine-tuned on synthetic RTL data.
• Performance metrics show Nemotron 3 Ultra achieves 97.1% average pass rate across CVDP categories while using only 6,629 tokens per iteration—28% fewer than GLM 5.2 and 71% fewer than Kimi K2.6.

Industry Insight

The success of Nemotron 3 Ultra in RTL coding suggests that specialized hybrid architectures trained on domain-specific synthetic data will become increasingly important for industrial AI applications, particularly in hardware design where precision and efficiency are critical. EDA tool vendors should prioritize integration with agentic AI systems that can leverage long-context reasoning and iterative feedback loops to accelerate chip design workflows.

TL;DR

• ACE-RTL 代理结合 NVIDIA Nemotron 3 Ultra 在 RTL 工作流中实现了最先进的准确性和效率，通过迭代生成-测试-反思方法和长上下文、RTL 专用推理。
• 在 CVDP 基准测试中，ACE-RTL 与 Nemotron 3 Ultra 在九类 RTL 任务中平均通过率达到 97.1%，比 GLM 5.2 和 Kimi K2.6 更高效，每迭代使用 Token 数减少高达 71%。
• Nemotron 3 Ultra 采用混合 Mamba-Attention 混合专家架构，基于多样化、规则过滤的合成 RTL 数据集训练，支持高吞吐、低成本的长上下文推理，便于与 EDA 工具集成。
• 现代芯片设计中，RTL 开发和验证需要精确推理和反复与 EDA 工具交互，AI 代理通过验证反馈迭代修正错误，显著提升设计效率。
• ACE-RTL 代理通过生成器、反思器和协调器三组件协作，实现迭代式上下文演化，大幅提升了模型在复杂 RTL 任务中的表现。

为什么值得看

该研究展示了 AI 代理与专用大模型在硬件设计领域的深度融合，为 RTL 开发提供了可落地的自动化解决方案。其高效推理与高准确率结合，显著降低芯片设计周期与成本，对半导体行业 AI 化转型具有战略意义。

技术解析

• ACE-RTL 代理采用“生成-测试-反思”迭代工作流，包含生成器（生成/更新 RTL 代码）、反思器（分析仿真反馈并定位根因）、协调器（维护上下文演化历史），实现闭环调试优化。
• 模型在 CVDP 基准测试中覆盖 9 类真实 RTL 任务，包括代码补全、自然语言转代码、模块复用、调试修复等，全面评估模型在工程场景下的泛化能力。
• Nemotron 3 Ultra 为 550B 总参数、55B 活跃参数的混合 Mamba-Attention 混合专家模型，支持 1M-token 上下文，基于 20T 文本 Token 预训练，并针对 RTL 任务进行专项微调。
• 在相同代理框架下，Nemotron 3 Ultra 平均每迭代仅使用 6,629 个 Token，比 GLM 5.2 少 28%，比 Kimi K2.6 少 71%，同时保持最高 97.1% 的平均通过率，体现卓越能效比。
• 实验表明，引入 ACE-RTL 代理后，各模型在调试类任务（cid016）通过率显著提升，如 GLM 5.2 从 44.0% 提升至 94.3%，Kimi K2.6 从 68.6% 提升至 97.1%，Nemotron 3 Ultra 达到 100%。

行业启示

• AI 代理与专用大模型的结合将成为硬件设计自动化的核心范式，推动 EDA 工具链向“智能协同”演进，建议厂商加速集成此类能力以提升设计效率。
• 长上下文与低 Token 消耗的模型架构是支撑复杂工程任务的关键，未来研发应聚焦于在保持推理精度的前提下进一步优化计算与内存效率。
• 合成数据与领域对齐训练在垂直领域（如 RTL 开发）中表现突出，企业应构建高质量、规则过滤的领域数据集，以提升模型在实际工程中的可用性与可靠性。