Oracle announced a further expansion of its partnership with NVIDIA to run strategic NVIDIA AI applications on the new Oracle Cloud Infrastructure Remote Software Service (OCI) Supercluster. NVIDIA has selected OCI as the enterprise’s hyperscale cloud technology provider for its large-scale AI supercomputing service, NVIDIA DGX Cloud, and is running its generative AI cloud service, NVIDIA AI Foundations, which is delivered through OCI-based DGX Cloud.
NVIDIA DGX Cloud and the NVIDIA AI Foundations services running on it leverage the NVIDIA certified OCI Supercluster to ensure they meet the high standards of DGX Cloud.
The OCI Supercluster includes OCI Compute Bare Metal, an ultra-low latency RoCE cluster based on the NVIDIA network, and an option for HPC storage. NVIDIA has deployed and validated the supercluster to support thousands of OCI Compute Bare Metal instances to efficiently process massively parallel applications. the OCI Supercluster network can now scale to 4,096 OCI Compute Bare Metal instances using 32,768 A100 GPUs. OCI Compute Bare Metal instances with NVIDIA H100 GPUs are now available in limited quantities.
In addition, NVIDIA announced that Oracle is adding NVIDIA BlueField-3 DPUs to its networking stack.
NVIDIA AI Foundations modeling services span language, image, video and 3D, as well as biology. Enterprises can use NVIDIA NeMo Language Services and NVIDIA Picasso Image, Video and 3D Services to build proprietary, domain-specific generative AI applications that enable intelligent chat and customer support, professional content creation, digital simulations and more. For bio-AI model training and inference, NVIDIA BioNeMo Cloud Services provides tools for rapid customization and deployment of generative AI applications.
On OCI, custom models using NVIDIA AI Foundations and model families like GPT-3 can benefit from the OCI Supercluster, which includes a purpose-built RDMA network. This network provides near wire-speed performance and microsecond latency and removes barriers related to RDMA-dependent workloads.