Microsoft deploys AMD Helios AI accelerators on Azure

The hunger for artificial intelligence compute is reaching fever pitch, fueled by the emergence of frontier-class open models that demand exponentially more processing power. As these massive systems demand unprecedented capacity, a critical infrastructure race is underway, with major players teaming up to provide the necessary FLOPS.

Microsoft and AMD have announced a significant strategic partnership aimed at supercharging AI capabilities within data centers. The goal is clear: to equip Redmond with the high-performance AI accelerators needed for both internal workloads and external customers utilizing Azure services.

The collaboration focuses on integrating AMD’s advanced Helios rack-scale AI accelerator into the Microsoft ecosystem. This move aims to make next-generation AI compute readily available for running demanding model training and inference serving tasks across the board.

For Microsoft, this partnership provides a powerful platform. It ensures that AI labs can access cutting-edge compute resources while also underpinning managed compute services offered through the Microsoft Foundry, supporting enterprise customers deploying AI workloads efficiently.

The Helios accelerator itself is packed with serious power. The system is built around 72 next-generation Instinct MI455X GPUs, boasting an impressive aggregate memory capacity of 31.1TB of HBM4. These components deliver formidable computational punch, offering up to 1.4 exaFLOPS in FP8 compute and 2.9 exaFLOPS in FP4 for AI models utilizing OCP AI data types.

Beyond raw processing power, the Helios architecture is designed for blistering speed. AMD is targeting a massive bandwidth capability within the rack, aiming for 260 TB/s of scale-up bandwidth. Furthermore, it supports 43 TB/s of scale-out bandwidth via UALink over Ethernet, positioning the system to handle data movement as quickly as the computation itself.

The ambition extends beyond just GPUs. The alliance also involves integrating AMD’s CPU innovations into Microsoft’s infrastructure. Azure will receive new Virtual Machine series built on AMD’s upcoming sixth-generation Epyc Venice CPUs. These processors are specifically tailored for demanding tasks, including agentic AI and complex data pipelines, alongside workflows for semiconductor design.

To fully realize this integration, Microsoft plans to leverage its existing deployment of AMD Pensando DPUs, further accelerating networking and storage processing operations across the Azure infrastructure. This holistic approach ensures that the hardware is not just powerful, but perfectly integrated into a cohesive system.

This strategic move represents a serious push in the data center GPU market. By delivering high-performance alternatives, AMD is poised to capture significant share from competitors as the AI compute demand continues to skyrocket globally. The partnership underscores the fact that the future of AI infrastructure relies on powerful, interconnected hardware designed for extreme scale and speed.