CoreWeave (CRWV.US) AI cloud computing power upgraded again! First to launch multi-rack Nvidia (NVDA.US) Vera Rubin NVL72 cluster

The Zhitong Finance App learned that CoreWeave (CRWV.US) announced that it has launched a multi-rack NVDA.US (NVDA.US) Vera Rubin NVL72 cluster on CoreWeave Cloud to connect hundreds of Nvidia Rubin graphics processors (GPUs) into a single system for intelligent AI (agentic AI) workloads. According to CoreWeave, this multi-rack architecture allows training and inference workloads to be scheduled and run across hundreds of Rubin GPUs, providing computing power support for larger models, higher intensity inference, and large-scale reinforcement learning.

A single Nvidia Vera Rubin NVL72 rack integrates 72 Rubin GPUs with 36 Vera CPUs, and is equipped with Nvidia's networking and data processing technology. CoreWeave said multiple racks are interconnected via Nvidia Spectrum-X Ethernet and operate together as a set of horizontally extended clusters. CoreWeave's infrastructure automates rack deployment, firmware upgrades, verification, and power and cooling, and performs full rack level testing before the rack is put into production.

Chen Goldberg, executive vice president and head of product and engineering at CoreWeave, said that as far as he knows, CoreWeave is the first AI cloud service provider to complete the Vera Rubin NVL72 verification and put it into operation. “With multi-rack Vera Rubin, we are connecting hundreds of Rubin GPUs into a single scale-out cluster,” he said. He added that the system can provide customers developing intelligent AI with a larger scale and faster iteration speed.

While expanding computing power, CoreWeave also introduced cross-region write acceleration for its AI object storage platform. This feature allows AI workloads to write data in the region where they are running, while CoreWeave copies data to another region in the background. The company says this feature can reduce downtime for customers running training workloads in multiple regions.

The AI cloud computing company also launched the Archive storage layer, providing a lower cost storage option for customers looking to keep their data for a long time. This storage tier does not charge for data retrieval, early deletion, or reading data from the Archive storage tier. Its native object transfer accelerator, LOTA, can also cache data closer to AI workloads. CoreWeave said the feature reduces read latency by 8 times compared to reading data from traditional storage clusters.