NVIDIA Nemotron 3 Ultra Leads Open Models in Agentic RTL Accuracy, Efficiency
NVIDIA Nemotron 3 Ultra has emerged as the leading open model for agentic register transfer level chip design, demonstrating superior accuracy and computational efficiency in complex hardware development workflows. Traditional RTL engineering is constrained by extensive verification cycles requiring precise temporal reasoning and iterative interaction with electronic design automation tools. To address this bottleneck, NVIDIA combined its new large language model with the ACE-RTL agent framework, which employs an automated generate-test-reflect pipeline comprising generator, reflector, and coordinator components to continuously refine code based on simulation feedback. The integrated system was evaluated using the Comprehensive Verilog Design Problems benchmark, which tests realistic RTL generation, modification, debugging, and verification tasks. Under controlled testing with the ACE-RTL agent, Nemotron 3 Ultra achieved a 97.1 percent average pass rate across nine task categories, surpassing competitors including GLM 5.2 and Kimi K2.6. The model also reached perfect scores on multiple debugging and code improvement categories. Crucially, Nemotron 3 Ultra maintains this performance with significantly reduced computational overhead, averaging just 6,629 tokens per iteration. This represents a 28 percent reduction compared to GLM 5.2 and a 71 percent reduction versus Kimi K2.6, directly translating to faster inference cycles and lower operational costs for engineering teams. The model capabilities stem from a 550 billion total-parameter mixture-of-experts architecture featuring 55 billion active parameters and a hybrid Mamba-Attention design. This configuration optimizes memory footprint and decoding speed, enabling stable processing of extended one-million-token contexts essential for long-running agentic debugging sessions. Training data was engineered through a specialized synthetic generation pipeline that moves beyond static specification-to-code translation. By injecting realistic hardware faults, simulating timing violations, and generating structured debugging datasets, NVIDIA ensured the model learns to interpret failure states and execute targeted code repairs rather than relying solely on forward generation. Industry adoption is already accelerating through strategic EDA partnerships. Cadence, Siemens, and Synopsys have integrated Nemotron 3 Ultra into their respective autonomous verification and design suites. These integrations enable multi-agent orchestration across front-end RTL development, testbench creation, and regression management, promising to compress weeks of manual verification into significantly shorter development windows. The release establishes a new baseline for open-source artificial intelligence in hardware engineering, offering semiconductor firms a scalable pathway to automate complex chip design workflows while maintaining strict accuracy standards and resource efficiency.
