Oracle and AMD have at the moment introduced a serious extension of their lengthy‑standing, multi‑technology partnership aimed toward serving to clients scale their AI capabilities considerably. From calendar Q3 2026, Oracle will develop into the primary hyperscaler to supply a publicly accessible AI supercluster powered by 50,000 AMD Intuition MI450 Collection GPUs, with additional enlargement deliberate for 2027 and past.
The transfer builds on a number of years of joint improvement, starting with the launch of AMD Intuition MI300X shapes on Oracle Cloud Infrastructure (OCI) in 2024 and the next rollout of OCI Compute with AMD Intuition MI355X GPUs. These choices feed into the broader zettascale OCI Supercluster.
The brand new AI superclusters will make use of AMD’s “Helios” rack design, integrating AMD Intuition MI450 GPUs, the upcoming “Venice” EPYC CPUs, and Pensando superior networking codenamed “Vulcano.” This vertically optimised rack‑scale structure is designed to ship excessive efficiency, vitality effectivity and scalability for giant‑scale AI coaching and inference workloads.
Mahesh Thiagarajan, Govt Vice President, Oracle Cloud Infrastructure, stated: “Our clients are constructing a number of the world’s most bold AI functions, and that requires strong, scalable, and excessive‑efficiency infrastructure. By bringing collectively the most recent AMD processor improvements with OCI’s safe, versatile platform and superior networking powered by Oracle Acceleron, clients can push the boundaries with confidence. By means of our decade‑lengthy collaboration with AMD—from EPYC to AMD Intuition accelerators—we’re persevering with to ship the very best value‑efficiency, open, safe, and scalable cloud basis in partnership with AMD to fulfill buyer wants for this subsequent period of AI.”
Forrest Norrod, Govt Vice President and Normal Supervisor of AMD’s Knowledge Heart Options Enterprise Group, added: “AMD and Oracle proceed to set the tempo for AI innovation within the cloud. With our AMD Intuition GPUs, EPYC CPUs, and superior AMD Pensando networking, Oracle clients achieve highly effective new capabilities for coaching, wonderful‑tuning, and deploying the subsequent technology of AI. Collectively, AMD and Oracle are accelerating AI with open, optimised, and safe techniques constructed for large AI information centres.”
The MI450 GPU‑powered shapes are engineered to assist refined workloads together with generative AI and excessive‑efficiency computing. Every GPU delivers as much as 432 GB of HBM4 reminiscence and 20 TB/s of reminiscence bandwidth, enabling customers to coach and infer fashions 50 % bigger totally in‑reminiscence. The “Helios” rack design permits dense, liquid‑cooled, 72‑GPU racks and combines UALoE scale‑up connectivity with Ethernet‑based mostly scale‑out networking for low latency and excessive throughput throughout GPU clusters.
As well as, the supercluster structure options superior head nodes based mostly on EPYC “Venice” CPUs providing confidential computing and security measures. AMD Pensando DPUs allow converged networking that gives line‑fee information ingestion, improves efficiency, and enhances safety. The design permits GPUs to be outfitted with as much as three 800 Gbps AI‑NICs (Vulcano), supporting lossless, excessive‑pace programmable connectivity with requirements akin to RoCE and UEC.
The structure additionally incorporates the open UALink and UALoE cloth, which simplifies information sharing throughout GPUs inside racks with out routing by means of CPUs. The open‑supply ROCm software program stack facilitates vendor flexibility and helps well-liked AI and HPC frameworks, whereas superior partitioning and virtualization options allow safe multi‑tenant useful resource sharing and wonderful‑grained allocation of GPU workloads.
Alongside the MI450 deployment, Oracle additionally introduced the overall availability of OCI Compute cases powered by AMD Intuition MI355X GPUs. These shall be a part of the zettascale OCI Supercluster able to scaling to 131,072 GPUs, providing clients different optimisation and suppleness for AI workloads.
