Oracle and AMD Forge Ahead: A Deep Dive into the 2026 AI Supercluster Powered by AMD MI450 GPUs
Introduction: A New Era of AI Scale
The relentless advancement of artificial intelligence has created an unprecedented demand for computing power. As AI models grow in complexity and scale, existing infrastructure is being pushed to its limits. In response to this escalating need, Oracle and AMD have announced a significant expansion of their long-standing partnership. This collaboration is set to introduce a new generation of AI superclusters, with Oracle poised to be the first hyperscaler to offer a publicly accessible AI supercluster powered by an initial deployment of 50,000 AMD Instinct MI450 Series GPUs, slated for availability in calendar Q3 2026. This initiative represents a pivotal step towards enabling customers to achieve next-generation AI scale, offering a robust, flexible, and highly performant cloud foundation.
The AMD Instinct MI450 Series GPU: Powering the Future of AI
At the heart of this new AI supercluster lies the AMD Instinct MI450 Series GPU. These accelerators are engineered to deliver breakthrough compute and memory capabilities, essential for tackling the most demanding AI workloads. Each MI450 GPU boasts an impressive memory capacity of up to 432 GB of HBM4 and an astounding 20 TB/s of memory bandwidth. This substantial increase in memory bandwidth allows customers to train and infer models that are up to 50 percent larger than previous generations, all within memory. This capability significantly reduces the need for complex model partitioning and accelerates the time-to-results for intricate AI tasks, including advanced language models, generative AI, and high-performance computing (HPC) workloads.
AMD's "Helios" Rack Design: Optimized for Extreme Scale
The deployment of the MI450 GPUs will be housed within AMD's innovative "Helios" rack design. This vertically-optimized, rack-scale architecture is meticulously engineered to deliver maximum performance, scalability, and energy efficiency. The "Helios" design integrates dense, liquid-cooled configurations, enabling a higher concentration of GPUs within each rack, thereby optimizing performance density and reducing operational costs and energy consumption. This approach is crucial for hyperscale AI data centers, where efficiency and power management are paramount. The "Helios" rack design also incorporates UALoE scale-up connectivity and Ethernet-based Ultra Ethernet Consortium (UEC)-aligned scale-out networking, which are designed to minimize latency and maximize throughput across interconnected racks and pods.
Synergistic Components: EPYC CPUs and Pensando Networking
Complementing the power of the AMD Instinct MI450 Series GPUs are next-generation AMD EPYC™ CPUs, codenamed "Venice." These powerful processors will serve as the head nodes within the supercluster, designed to maximize cluster utilization and streamline large-scale workflows by accelerating job orchestration and data processing. Importantly, these EPYC CPUs will feature confidential computing capabilities and built-in security features, offering end-to-end protection for sensitive AI workloads. The networking infrastructure is equally advanced, leveraging next-generation AMD Pensando™ advanced networking codenamed "Vulcano." This DPU-accelerated converged networking solution is built on fully programmable AMD Pensando DPU technology, providing the high-performance and security required for data centers to handle the next era of AI training, inferencing, and cloud workloads. Each GPU can be equipped with up to three 800 Gbps AMD Pensando "Vulcano" AI-NICs, ensuring lossless, high-speed, and programmable connectivity that supports advanced RoCE and UEC standards, crucial for efficient distributed training and collective communication.
Interconnectivity and Software: UALink, ROCm, and Open Standards
The architecture is further enhanced by innovative UALink and UALoE fabric technologies. These facilitate efficient workload expansion, reduce memory bottlenecks, and enable the orchestration of massive multi-trillion-parameter models. UALink, an open, high-speed interconnect standard purpose-built for AI accelerators, minimizes hops and latency by allowing direct, hardware-coherent networking and memory sharing among GPUs within a rack, bypassing CPUs. This adherence to open standards ensures flexibility, scalability, and reliability for demanding AI workloads. Furthermore, the inclusion of the open-source AMD ROCm™ software stack provides customers with a flexible programming environment, including popular frameworks, libraries, compilers, and runtimes. This open approach fosters rapid innovation, offers freedom of vendor choice, and simplifies the migration of existing AI and HPC workloads.
OCI
AI Summary
The collaboration between Oracle and AMD marks a significant advancement in AI infrastructure, with Oracle set to be the first hyperscaler to offer a publicly available AI supercluster powered by 50,000 AMD Instinct MI450 Series GPUs, beginning in calendar Q3 2026. This initiative is a direct response to the rapidly growing demand for AI capacity, as next-generation AI models increasingly outgrow the capabilities of current AI clusters. The new superclusters will be built upon AMD's "Helios" rack design, integrating AMD Instinct MI450 Series GPUs, next-generation AMD EPYC CPUs codenamed "Venice," and advanced AMD Pensando networking codenamed "Vulcano." This vertically-optimized, rack-scale architecture is engineered for maximum performance, scalability, and energy efficiency, crucial for large-scale AI training and inference workloads. Oracle's commitment to this partnership, building on a decade of co-innovation, aims to deliver unparalleled price-performance, security, and scalability. The AMD Instinct MI450 Series GPUs themselves offer breakthrough compute and memory capabilities, with up to 432 GB of HBM4 and 20 TB/s of memory bandwidth, enabling customers to train and infer models up to 50 percent larger than previous generations entirely in-memory. The "Helios" rack design emphasizes dense, liquid-cooled configurations for optimal performance density and energy efficiency. Complementing the GPUs are the "Venice" EPYC CPUs, providing powerful head node capabilities for job orchestration and data processing, alongside confidential computing features. The advanced networking, powered by AMD Pensando DPUs and "Vulcano" AI-NICs, ensures high-speed, low-latency connectivity essential for distributed training and collective communication. Key interconnect technologies like UALink and UALoE are integrated to minimize latency and memory bottlenecks for massive models. The open-source AMD ROCm software stack further empowers customers with flexibility and ease of migration. Additionally, advanced partitioning and virtualization features will allow for safe and efficient resource sharing. This strategic deployment by Oracle, leveraging AMD's cutting-edge hardware and OCI's robust cloud platform, is poised to redefine the landscape of AI infrastructure, offering a powerful, open, and secure foundation for the future of artificial intelligence.