Nvidia has formally launched the Vera Rubin platform, a combination CPU and GPU platform billed as a major step forward in the convergence of artificial intelligence and high-performance computing (HPC) for scientific research. Announced at the ISC High Performance 2026 conference in Hamburg, the new platform combines Nvidia’s Vera CPUs, Rubin GPUs, networking technologies, and software stack into what the company describes as a rack-scale supercomputer. Nvidia is targeting scientific workloads, from climate modeling and computational fluid dynamics to quantum chemistry, energy exploration, and large data center operators.

“Nvidia’s roots are firmly planted in scientific computing, and native FP64 precision remains absolutely vital for accurate fluid dynamics, climate modeling, and geoscience,” said Dion Harris, senior director of HPC and AI factory solutions at Nvidia, on a conference call. “We are committed to maintaining that support moving forward.”

At the heart of the platform is a tightly integrated architecture combining Nvidia Rubin GPUs and Vera CPUs linked through NVLink-C2C interconnects, ConnectX-9 SuperNICs, and BlueField-4 DPUs. The systems are built around direct liquid cooling and support up to 144 GPUs in a single rack.

Nvidia claims a fully configured Vera Rubin system can deliver more than seven exaflops of AI performance for scientific workloads alongside five petaflops of native double-precision (FP64) computing performance. That would put a Rubin system well ahead of the top supercomputers in the TOP500 ranking. Updated rankings are due later this week.

The Vera Rubin architecture increases memory bandwidth by 2.8 times compared to Blackwell, the previous generation GPU. “We are projecting up to four times performance boosts for memory-bound fluid dynamic applications,” said Harris. “With Rubin, we are ensuring that the fundamental mathematical workloads driving scientific discovery run faster, more efficiently, and with greater precision than ever before.”

The new platform is designed to support both traditional HPC simulations and emerging AI-driven scientific applications. Researchers will be able to train foundation models, deploy surrogate models, run simulations and perform real-time data analysis on a single infrastructure.

“AI is shifting from a tool that simply answers questions to an autonomous system that executes complex tasks,” said Harris. “Early data shows [agentic AI] increases simulation demand by up to ten times.”

Nvidia also announced that several leading research institutions announced plans to build next-generation systems based on the new architecture. The Leibniz Supercomputing Centre (LRZ) in Germany will deploy Vera Rubin in its upcoming Blue Lion supercomputer, scheduled to enter service in 2027.

Blue Lion is a second-generation exascale-class HPE Cray system, and it is expected to deliver approximately 30 times the computing power of LRZ’s current system, supporting research in astrophysics, environmental science and life sciences.

In the U.S., the National Energy Research Scientific Computing Center (NERSC) will use Vera Rubin technology in Doudna, the next flagship supercomputer for the Department of Energy at Lawrence Berkeley National Laboratory. This system is being built by Dell Technologies and will support large-scale HPC simulations, AI training and data-intensive research.

Meanwhile, Los Alamos National Laboratory has selected Vera Rubin technology for three new supercomputers: Mission, Vision and Veritas. Mission will focus on national security workloads, while Vision will support open scientific research and AI-driven discovery. Veritas is specifically designed to enable agentic AI applications in scientific research, combining Rubin GPUs with standalone Vera CPU partitions.

Vera Rubin NVL4-based systems from Dell and Super Micro were also announced at the event.

Background and context: The evolution of Nvidia's HPC strategy

Nvidia’s journey in high-performance computing began more than a decade ago with the introduction of CUDA, which allowed GPUs to be used for general-purpose computing. Over successive generations, from Kepler to Volta, Turing, Ampere, Hopper, and Blackwell, Nvidia steadily increased both raw compute power and memory bandwidth. The Rubin platform represents a significant architectural leap, integrating CPU and GPU cores in a unified memory space through NVLink-C2C. This reduces data movement bottlenecks and simplifies programming. The decision to include native FP64 support is notable because some earlier HPC-focused GPUs, such as the A100 and H100, offered reduced FP64 throughput compared to their tensor core capabilities. Rubin restores full double-precision performance, catering to scientific codes that require high accuracy.

Another key innovation is the adoption of direct liquid cooling at the rack level. Modern GPUs generate substantial heat, especially when comprising 144 units per rack. Liquid cooling improves efficiency and reliability, allowing denser deployment without thermal throttling. This aligns with broader industry trends toward immersion and cold-plate cooling in data centers.

Technical deep dive: Vera CPU and Rubin GPU

The Vera CPU is not based on x86 but on Nvidia’s custom Arm architecture, specifically the “Grace” line of server processors first introduced in 2023. The Vera CPU features up to 144 cores and support for next-generation memory standards. It is designed to complement the Rubin GPU, which succeeds the Blackwell architecture. Rubin GPU leverages advanced node technology (likely N3P or similar) and incorporates HBM4 memory for bandwidth exceeding 5 TB/s. The combination of Vera and Rubin via NVLink-C2C provides cache coherence and high-speed interconnects, essential for workloads that require frequent CPU-GPU communication.

The memory bandwidth increase—2.8 times over Blackwell—is dramatic. For memory-bound applications like computational fluid dynamics (CFD), this directly translates to faster time-to-solution. Nvidia’s projection of up to 4x performance boost for fluid dynamics is plausible given the bandwidth scaling coupled with improved parallel processing capabilities. The FP64 performance boost also enables legacy HPC codes to run without modification, ensuring compatibility.

Market implications and competitive landscape

Nvidia’s announcement places pressure on competitors like AMD, Intel, and emerging custom silicon from hyperscalers. AMD’s Instinct MI400 series and Intel’s Falcon Shores are currently in development, but none have demonstrated exascale-class FP64 performance with AI capabilities in a single rack. The Vera Rubin platform could become the de facto standard for supercomputing centers planning next-generation systems. However, cost and power consumption remain challenges: a full rack with 144 GPUs likely draws hundreds of kilowatts. Nvidia’s liquid cooling solution mitigates thermal issues but requires infrastructure upgrades.

Research institutions often have multi-year procurement cycles. By securing early adopters like LRZ, NERSC, and Los Alamos, Nvidia gains credibility and validation. The Blue Lion system in Germany will provide a high-profile showcase by 2027. Similarly, Doudna at NERSC will demonstrate integration with Dell’s enterprise offerings. The separation of Mission, Vision, and Veritas at Los Alamos shows how the platform can address diverse workflows—national security, open science, and autonomous AI.

Agentic AI and the future of scientific simulation

The concept of agentic AI—autonomous systems that plan and execute tasks—is still nascent but holds promise for accelerating discovery. In scientific simulation, agentic AI could autonomously adjust model parameters, run sensitivity studies, and propose experiments. This is computationally intensive, requiring both traditional HPC for simulation and AI for inference and reasoning. Rubin’s unified architecture supports both within the same node, avoiding data transfer latencies. Nvidia’s claim of up to 10x increase in simulation demand for agentic AI suggests that once such capabilities are deployed, compute requirements will skyrocket, further driving the need for platforms like Vera Rubin.

Overall, the Vera Rubin platform represents Nvidia’s most ambitious integration of CPU, GPU, networking, and cooling to date. It is poised to dominate the supercomputing sector for the next several years, especially as the lines between HPC and AI blur. The announcements at ISC 2026 underscore Nvidia’s commitment to scientific computing and its ability to deliver hardware that meets the highest standards of precision and performance.

Source: Network World News