NVIDIA Vera Rubin at ISC 2026: Specs, Deployments, and the Convergence of HPC and AI

At ISC High Performance 2026 in Hamburg, Germany, NVIDIA unveiled the Vera Rubin accelerated computing platform, purpose-built for scientific supercomputing. It combines NVIDIA Rubin GPUs and Vera CPUs connected via NVLink-C2C in a full-stack system designed for both traditional simulation and AI workloads .

A single rack of Vera Rubin can deliver over 7 exaflops of AI performance and 5 petaflops of native double-precision (FP64) computing, using up to 144 GPUs — performance comparable to systems on the TOP500 list . The Vera CPU features 88 cores with 2x performance-per-watt versus previous-generation NVIDIA CPUs .

Key Performance Specifications

• Native FP64 double-precision performance for traditional HPC simulation workloads, delivered through the Rubin GPU architecture .
• NVL4 configuration: Supermicro's DCBBS Blueprint based on Vera Rubin NVL4 supports up to 1,152 NVIDIA Rubin GPUs and 576 NVIDIA Vera CPUs in a single scalable unit .
• 3.2 MW scalable unit with DLC-2 Direct Liquid Cooling supporting 362 kW per rack .
• The Rubin platform (first announced at CES 2026) delivers up to 10x reduction in inference token cost and 4x reduction in GPU count for training mixture-of-experts models versus the Blackwell platform .
• Integrated with NVIDIA CUDA-X libraries, NVIDIA NIM microservices, NVIDIA AI Enterprise software, and NVIDIA Quantum InfiniBand networking .

Research Institutions Deploying Vera Rubin Systems

A record 35 new NVIDIA AI HPC supercomputers are in development across 23 European countries, equipping more than 3 million researchers . Named institutions include:

• Barcelona Supercomputing Center — EuroHPC AI Factory
• BavariaAI — Blue Swan system
• IT4Innovations (IT4I) — Czech Republic
• HLRS (High-Performance Computing Center Stuttgart) — HammerHAI system
• NAISS — Mimer EuroHPC AI Factory, Sweden
• CINECA (Italy)
• Fraunhofer FOKUS (Germany)
• Jülich Supercomputing Centre (Forschungszentrum Jülich)

These systems span AI factories, climate and energy research, and quantum-GPU hybrid computing .

Industry Partnerships and Manufacturing

• Supermicro: Delivered the first end-to-end DCBBS Blueprint for HPC based on the NVIDIA Vera Rubin NVL4 platform, providing liquid-cooled racks and complete integration .
• Corning: Multiyear partnership to expand U.S.-based manufacturing of advanced optical connectivity solutions for AI infrastructure, including three new facilities in North Carolina and Texas .
• Industrial software leaders (Cadence, Dassault Systèmes, PTC, Siemens, Synopsys): Bringing CUDA-X, Omniverse, and GPU-accelerated tools to manufacturing and engineering workflows .
• Siemens: Expanded partnership to build an AI-accelerated industrial operating system spanning design, engineering, manufacturing, and supply chains .

AI Integration with Traditional Scientific Simulation

Vera Rubin is designed as a converged HPC and AI platform — running both traditional physics-based simulation (FP64) and AI workloads on the same infrastructure .

• PhysicsNeMo: NVIDIA's AI-physics framework for CFD, thermal analysis, climate, and weather modeling is supported on Vera Rubin. It uses machine learning surrogates to accelerate multiscale models and traditional solvers to generate training data for AI .
• Climate modeling: Vera Rubin accelerates climate simulation by combining native FP64 solvers with AI-driven emulation and data assimilation .
• Computational fluid dynamics (CFD): The platform enables AI-accelerated CFD where neural surrogates speed up iterative solver loops while maintaining double-precision accuracy for validation .
• Energy exploration: Seismic imaging, reservoir simulation, and renewable energy modeling benefit from AI proposing candidate scenarios while traditional FP64 solvers validate them .

NVIDIA describes this convergence as science living "at the intersection of physics-based simulation and AI — where machine learning surrogates accelerate multiscale models, generative AI proposes candidate molecules or materials for high-fidelity validation, and traditional solvers generate the training data that makes AI possible" .