NVIDIA Blackwell dominates MLPerf, launches Nemotron 3 Ultra with 5x speed boost
NVIDIA's Blackwell sets records, cuts AI costs 30%, and reindustrializes US with Texas fab.
NVIDIA announced multiple breakthroughs across AI hardware and software at its latest event. The Blackwell GPU architecture swept the MLPerf Training 6.0 benchmark, proving itself as the fastest, largest, and strongest platform for training frontier models like GPT-5 and Llama 4. This performance win underscores NVIDIA's dominance in enterprise AI infrastructure, enabling customers to train models faster and at scale.
In addition to hardware wins, NVIDIA launched Nemotron 3 Ultra, a frontier-reasoning model designed for long-running agentic tasks. It delivers 5x faster inference than previous Nemotron models and up to 30% lower cost compared to open frontier models like Llama 3. The GB300 NVL72 also topped the new Artificial Analysis AgentPerf benchmark, running 20x more AI agents per megawatt, making it ideal for always-on enterprise assistants. On the manufacturing front, NVIDIA partnered with Coherent to expand a Texas factory that will produce 6-inch indium phosphide (InP) photonics, a critical material for next-gen high-speed interconnects in AI factories. This move aims to reindustrialize the US and reduce reliance on overseas supply chains.
- NVIDIA Blackwell leads MLPerf Training 6.0 with record performance across all workloads.
- Nemotron 3 Ultra offers 5x faster inference and cuts costs up to 30% vs. open frontier models.
- Texas factory with Coherent will produce first 6-inch InP photonics to bolster US AI manufacturing.
Why It Matters
NVIDIA's hardware and software advances accelerate AI deployment while reshoring critical manufacturing, lowering costs for enterprises.