Oracle to launch AI supercluster with AMD MI450 GPUs in 2026

Oracle and AMD have announced an expansion of their partnership to support customers in scaling their artificial intelligence initiatives.

Oracle is set to become the first hyperscaler to offer a publicly available AI supercluster powered by AMD Instinct MI450 Series GPUs. This new supercluster is planned for initial deployment of 50,000 GPUs in the third quarter of 2026, with an intention to expand further in subsequent years.

The collaboration draws on an established relationship between Oracle and AMD, including prior integration of AMD Instinct GPU platforms into Oracle Cloud Infrastructure (OCI). Previous offerings began with the rollout of AMD Instinct MI300X-powered shapes and have extended to include the general availability of OCI Compute scenarios using AMI355X GPUs within the zettascale OCI Supercluster.

The announcement follows growing demand for large-scale AI infrastructure as advanced models require more sophisticated computational resources. OCI's upcoming superclusters will use the AMD "Helios" rack design, incorporating the latest AMD Instinct MI450 Series GPU with next-generation AMD EPYC CPUs and AMD Pensando advanced networking. 

"By bringing together the latest AMD processor innovations with OCI's secure, flexible platform and advanced networking powered by Oracle Acceleron, customers can push the boundaries with confidence," said Mahesh Thiagarajan, Executive Vice President, Oracle Cloud Infrastructure. "Through our decade-long collaboration with AMD-from EPYC to AMD Instinct accelerators-we're continuing to deliver the best price-performance, open, secure, and scalable cloud foundation in partnership with AMD to meet customer needs for this next era of AI."

The AMD Instinct MI450 Series GPU-powered shapes integrated within OCI are designed for high-performance and adaptable cloud deployment environments, supporting numerous open-source frameworks. This provision is intended to serve organisations running modern language models, generative AI, and high-performance computing workloads.

The AMD Instinct MI450 Series promises upgrades in compute and memory, offering up to 432GB of HBM4 memory and 20TB/s of bandwidth per GPU. This means AI models can be trained and inferred entirely in-memory, even when they are up to 50 per cent larger than those supported by the previous GPU generation.

The "Helios" rack design, utilising dense, liquid-cooled, 72-GPU racks, brings performance density and energy efficiency into focus. It employs UALoE scale-up connectivity and Ethernet-based scale-out networking compliant with Ultra Ethernet Consortium standards, aiming to reduce latency and maximise bandwidth across large clusters.

The cluster's head nodes will use AMD's next-generation EPYC CPUs, which introduce confidential computing features and built-in security measures intended to protect sensitive AI workloads throughout the computational process.

"With our AMD Instinct GPUs, EPYC CPUs, and advanced AMD Pensando networking, Oracle customers gain powerful new capabilities for training, fine-tuning, and deploying the next generation of AI," said Forrest Norrod, Executive Vice President and General Manager, Data Centre Solutions Business Group, AMD. "Together, AMD and Oracle are accelerating AI with open, optimized, and secure systems built for massive AI data centers."

The infrastructure incorporates UALink and UALoE fabric for efficient expansion, aiming to reduce memory transfer bottlenecks while providing direct, coherent networking and memory sharing amongst GPUs. UALink is promoted as an open, high-speed standard specifically designed for AI acceleration, and supported by a broad ecosystem.

The open-source AMD ROCm software stack will be provided for these systems, allowing flexibility in programming and simplifying the process of moving existing AI and high-performance computing workloads to the OCI platform. ROCm is compatible with prominent frameworks, libraries, compilers, and runtimes utilised by the AI research and development community.

Advanced partitioning and virtualisation features are available, allowing customers to securely and efficiently allocate resources, including granular GPU and pod partitioning, SR-IOV virtualisation, and multi-tenancy arrangements.

In parallel, OCI Compute with AMD Instinct MI355X GPUs has reached general availability, now accessible within the OCI Supercluster, which has capacity to scale to 131,072 GPUs. The MI355X-powered scenarios place emphasis on value, flexibility, and open-source software compatibility.