NVIDIA's Nemotron 3 Ultra Revolutionizes RTL Coding with Unmatched Efficiency

NVIDIA has made significant strides in artificial intelligence (AI) with its Nemotron 3 Ultra model, achieving a 100% pass rate on CVDP benchmarks. This achievement underscores its dominance in register transfer level (RTL) coding workflows, a critical area for modern chip design.

The Nemotron 3 Ultra is a 550-billion-parameter Mixture-of-Experts (MoE) model built for long-running, agentic AI tasks. It leverages a Hybrid Mamba-Attention architecture optimized for long-context reasoning, enabling it to handle the iterative nature of RTL coding. According to NVIDIA, the model achieves up to 5.9x higher inference throughput compared to competing models under specific token settings.

Nemotron 3 Ultra's performance on the CVDP benchmark is impressive, with a 97.1% average pass rate across nine categories. It outperformed other open models, including Kimi K2.6 and GLM 5.2, which scored 95.2% and 92.1%, respectively.

The model's efficiency is another standout feature, using 28% fewer tokens than GLM 5.2 and 71% fewer than Kimi K2.6. This makes it significantly more cost-effective for long-running workflows.