Oracle Picks AMD to Power New AI Supercluster With 50,000 GPUs
Oracle and AMD have worked together for several years to integrate AMD hardware into OCI, beginning with Instinct MI300X GPUs in 2024.
Topics
News
- Nexus Closes Eighth Fund at $700 Million for Early-Stage Bets in India and US
- Amazon to Pump $12.7 Billion into India’s AI and Cloud Expansion
- Anthropic Debuts Interviewer Tool to Understand How People Really Use AI
- IPL Tops Google’s Search List for 2025 as AI Breaks Into India’s Top 10
- TeamLease Names Suparna Mitra MD & CEO in Leadership Overhaul
- Apple Loses Top UI Designer as Alan Dye Joins Meta
[Image source: Chetan Jha/MITSMR India]
Oracle Corp. and Advanced Micro Devices Inc. (AMD) said on Tuesday they will jointly launch a new AI supercluster in 2026, deepening their long-standing partnership as demand surges for high-performance computing.
Oracle Cloud Infrastructure (OCI) will be the first to deploy AMD’s next-generation Instinct MI450 GPUs at scale.
The supercluster will go live with 50,000 GPUs in the third quarter of 2026, with plans for expansion in subsequent years.
The system is aimed at handling increasingly complex generative AI workloads and will use AMD’s new “Helios” rack architecture, combining MI450 GPUs, Venice EPYC CPUs, and Pensando networking components.
The companies said the setup is designed to improve performance, scalability, and energy efficiency for large-scale AI training and inference tasks.
Oracle and AMD have worked together for several years to integrate AMD hardware into OCI, beginning with Instinct MI300X GPUs in 2024.
The upcoming MI355X GPUs will be available soon on OCI’s zettascale supercluster.
The MI450 chips will feature 432 GB of HBM4 memory and 20 TB/s of bandwidth per GPU.
The racks will use a liquid-cooled, 72-GPU configuration connected via open UALink fabric to boost throughput and reduce latency.
The new system will also offer confidential computing, DPU-accelerated networking, and support for AMD’s ROCm software stack to ensure compatibility with widely used AI frameworks.
The announcement follows Oracle’s July agreement with OpenAI to deliver up to 4.5 gigawatts of additional “Stargate” capacity, part of a five-year deal reportedly worth more than $300 billion.
