Nvidia Unveils Vera Rubin Superchip with Next-Generation AI and HPC Integration

Nvidia has officially revealed its groundbreaking Vera Rubin Superchip at GTC 2025, marking a major leap in artificial intelligence and high-performance computing technology. This compact yet powerful design merges two Rubin GPUs with an 88-core Vera CPU, creating a unified platform optimized for unparalleled computational performance. According to Nvidia, production of all components is set to begin next year.

During the announcement, CEO Jensen Huang highlighted that the Vera Rubin architecture will deliver up to 100 PetaFLOPS FP4 computational performance, making it one of the most advanced systems ever developed. Early test units produced by TSMC are now being evaluated at Nvidia’s laboratories. Unlike traditional chips, these Superchips function more like fully equipped computing nodes that integrate CPU and GPU capabilities onto a single board, featuring high-bandwidth interconnects and onboard memory.

Revolutionary System Design

The Vera Rubin Superchip continues Nvidia’s tradition of merging general-purpose processing with extreme graphics performance. Each unit houses the Vera CPU built on 88 high-efficiency ARM-based cores with 176 threads, paired with two Rubin GPUs that integrate eight HBM4 memory stacks each. These components are covered by advanced heat spreaders built from precision-engineered aluminum for energy efficiency and improved thermal balance.

This integrated approach allows the Superchip to sustain exceptional bandwidth and deliver outstanding efficiency for deep learning and scientific workloads. The entire platform has been optimized for seamless performance scaling, minimizing latency and power loss even under intense workloads.

Next-Generation Performance Platforms

The first release, known as the Nvidia Vera Rubin NVL144, will pair two Rubin GPUs offering 50 PFLOPS FP4 capability and 288 GB of next-gen HBM4 memory. Driven by the Vera CPU, which links through a 1.8 TB/s NVLINK-C2C system, this model enhances throughput while maintaining low communication latency between components.

This architecture introduces a massive generational improvement. Compared to the previous GB300 NVL72, the NVL144 platform provides a 3.6 Exaflop FP4 inference speed and 1.2 Exaflop FP8 training capability, accompanied by 13 TB of HBM4 memory and 75 TB of high-speed buffer memory. Its NVLINK and CX9 interfaces are doubled, achieving 260 TB/s and 28.8 TB/s of throughput, respectively. The NVL144 system is projected to become commercially available in the latter half of 2026.

The next-generation step forward, the Nvidia Rubin Ultra NVL576, is scheduled for the second half of 2027. This massive expansion will scale the NVL structure to 576 processing units. The upgraded Rubin Ultra GPU will deliver 100 PFLOPS FP4 performance through a quad-GPU design and integrate 1 TB of HBM4e memory across 16 stacks.

Once deployed, the NVL576 will deliver 15 Exaflop FP4 inference and 5 Exaflop FP8 training throughput, featuring 4.6 PB of HBM4 memory and 365 TB of fast-access cache. These specifications represent over a tenfold increase in performance compared to the previous architecture. Additionally, connection bandwidth through NVLINK and CX9 has been expanded up to 1.5 PB/s and 115.2 TB/s, ensuring unmatched scalability for AI-driven research and data-intensive workloads.

The unveiling of the Vera Rubin Superchip solidifies Nvidia’s leadership in redefining the boundaries of computing by merging raw power, memory capacity, and intelligent architecture. With these advances, the company aims to pave a new path for AI systems of the future, enabling research units and enterprises to explore capabilities previously out of reach.

• #Nvidia Vera Rubin