NVIDIA Vera Rubin

Overview

NVIDIA Vera Rubin is a rack-scale computing platform developed to address the growing computational demands of artificial intelligence, particularly large language models, generative AI, and advanced reasoning systems. The platform combines multiple specialized processors and networking technologies into a unified architecture designed to maximize computing performance, memory efficiency, and communication between devices.

At the center of the platform are NVIDIA Rubin GPUs, which provide the parallel processing capabilities required for demanding AI workloads, and NVIDIA Vera CPUs, which coordinate data processing and support system operations. Based on NVIDIA's announced specifications, Rubin GPUs feature advanced HBM4 memory, offering up to 288 GB of high-bandwidth memory per GPU and about 22 TB/s of memory bandwidth, compared with 8 TB/s on the B300.

The NVIDIA Vera Rubin NVL72 configuration integrates 72 Rubin GPUs and 36 Vera CPUs within a rack-scale system, enabling large AI workloads to operate across tightly interconnected computing resources. By combining processing power, memory capacity, and high-speed networking, Vera Rubin is designed to support the increasing scale and complexity of modern AI infrastructure. Source: NVIDIA Developer

How it works

NVIDIA Vera Rubin operates through the coordinated execution of AI workloads across specialized computing components. The Rubin GPUs perform the intensive mathematical operations required for neural network training and inference, while Vera CPUs manage supporting computational tasks, data processing, and system-level operations. Each Rubin GPU uses high-bandwidth HBM4 memory to store model parameters, activations, and intermediate data, allowing rapid access to the information required during computation.

NVIDIA NVLink 6 provides high-speed communication between GPUs, delivering up to 3.6 TB/s of bidirectional bandwidth per GPU. This enables multiple GPUs to exchange information efficiently when processing large, distributed AI models. Additional networking technologies, including NVIDIA ConnectX-9 and Spectrum-6 Ethernet, support communication across servers and larger computing clusters.

Orchestration and infrastructure management software coordinate workload scheduling, resource allocation, and system monitoring to maintain efficient operation. Together, these technologies allow Vera Rubin systems to support large-scale distributed training, high-throughput inference, and complex AI applications requiring substantial computing resources.

Why it matters

NVIDIA Vera Rubin represents an important advancement in AI infrastructure because it addresses three major challenges in modern artificial intelligence: computational performance, memory limitations, and communication between GPUs. As AI models become larger and reasoning workloads more demanding, organizations require increasingly powerful computing systems capable of processing extensive datasets and generating large volumes of tokens efficiently.

Vera Rubin's advanced GPU architecture, HBM4 memory, and NVLink 6 interconnects are designed to improve performance while reducing infrastructure bottlenecks. According to NVIDIA, the Rubin platform is designed to deliver up to a tenfold reduction in inference token costs compared with Blackwell-based systems under its specified performance comparisons. Actual improvements depend on workload characteristics, system configuration, and operating conditions.

Understanding Vera Rubin is essential for organizations planning future GPU infrastructure investments, evaluating hardware generations, and preparing for increasingly demanding AI training and inference requirements.

Use cases