NVIDIA Unveils BlueField-4, Quantum Integration, and AI Physics Models at SC25

Details

• At SC25 in St. Louis, NVIDIA introduced a suite of AI infrastructure advancements, including the BlueField-4 data processing unit, quantum-GPU integration through NVQLink, and Apollo open models for AI-driven physics simulations.
• The BlueField-4 DPU merges a 64-core NVIDIA Grace CPU with the ConnectX-9 SuperNIC, providing 800Gb/s throughput and six times the compute power of its predecessor, BlueField-3, to offload networking, storage, and security tasks from main processors.
• NVQLink establishes a universal interconnect between quantum processors and NVIDIA GPU supercomputers, enabling efficient hybrid quantum-classical computing as shown in Quantinuum's Helios processor, which achieves real-time quantum error correction and ultra-low latency for scalable quantum-AI workflows.
• NVIDIA's collaboration with Japan’s RIKEN will deliver two new supercomputers powered by over 2,000 Blackwell GPUs for advanced research in life sciences, materials, climate modeling, and quantum science, targeting operational readiness in 2026.
• Top storage and infrastructure players including DDN, VAST Data, WEKA, TACC, Lambda, and CoreWeave are adopting BlueField-4 and Quantum-X Photonics switches, with WEKA projecting more than 100-fold efficiency gains in tokens-per-watt compared to traditional storage solutions.

Impact

NVIDIA’s innovations at SC25 showcase a tightly integrated approach to AI and quantum computing infrastructure, positioning the company at the forefront of next-generation scientific and enterprise workflows. By blurring the lines between quantum, AI, and classical computing with industry and research partnerships, NVIDIA is set to reshape computing paradigms and deepen its competitive lead in the era of AI supercomputing. Ongoing focus on efficiency and power management addresses critical scalability challenges as demand for intelligent infrastructure soars worldwide.