Oracle and Nvidia Partner to Deploy AI Supercluster for Advanced Reasoning Models

Oracle has deployed thousands of Nvidia GPUs across its Oracle Cloud Infrastructure (OCI) to develop and run next-generation AI reasoning models and agents. This marks the first wave of liquid-cooled NVIDIA GB200 NVL72 racks in OCI data centers, featuring thousands of Nvidia Grace CPUs and Blackwell GPUs.

Key Highlights

• NVIDIA GB200 NVL72: A supercomputer comprising 36 Arm-based Nvidia Grace CPUs, each paired with two Blackwell GPUs and connected via NVLink. Each GB200 NVL72 delivers over one exaflop of training performance.
• Scalability: Oracle plans to expand this into a cluster of over 100,000 Blackwell GPUs, forming one of its OCI Superclusters.
• Software Integration: The collaboration includes a full stack of software and database integrations.

Previous Achievements

Oracle previously built an OCI Supercluster with 65,536 Nvidia H200 GPUs using Hopper GPU technology, offering up to 260 exaflops of peak FP8 performance.

Availability

The Blackwell GPUs are accessible via:

• Oracle’s public, government, and sovereign clouds.
• Customer-owned data centers through OCI Dedicated Region and Alloy offerings. Learn more.

Competitive Landscape

Oracle joins other cloud providers like Google, CoreWeave, and Lambda in offering the GB200 NVL72 system. Microsoft also provides GB200 GPUs, though not configured as an NVL72 machine.

Technical Edge

The NVL72 stands out by presenting multiple CPUs and GPUs as a single image to the software, with a shared memory space. This contrasts with traditional setups where each GPU operates independently.

• NVLink 5th Gen: Enables 130 TB/s GPU-to-GPU interconnect bandwidth, facilitating rapid data sharing and synchronization—critical for training large AI models.
• Scalability Challenge: Most clusters struggle beyond eight GPUs, but NVL72’s design overcomes this barrier.