Microsoft Expands Azure AI Infrastructure With AMD

The ND MI455X v7 family will use AMD Helios for large-scale inference workloads, including reasoning, search and agent services. Helios combines Instinct MI455X GPUs, EPYC Venice CPUs, Pensando networking and AMD's ROCm software in an integrated rack-scale system.

The HDv2 family targets CPU-heavy data preparation, search, reinforcement learning and agent coordination. Microsoft lists nearly 500 physical sixth-generation EPYC cores, 4 TB of RAM, 32 TB of local NVMe storage and 400 Gb Azure Boost networking for the planned configuration.

The HXv2 family targets electronic-design automation and technical computing. Microsoft says it will offer 176 EPYC cores above 5 GHz, 50% more addressable cache per core, configurations with nearly 2 TB or 4 TB of RAM, and 800 Gb InfiniBand for distributed-memory workloads.

The agreement also broadens Microsoft's deployment of AMD Pensando data-processing units. AMD and Microsoft said they are integrating AMD silicon with Azure Boost to improve networking and connection processing across Azure infrastructure. That is important because moving data between storage, hosts and accelerators can limit usable inference throughput even when GPU capacity is available.

Tom's Hardware noted that the announcement does not quantify Microsoft's Helios purchase. It also places the deployment in AMD's broader effort to gain data-center accelerator share against Nvidia. Those competitive implications are interpretation; the verified product commitments are the upcoming Azure families, the component mix and AMD's second-half shipment target.

For practitioners, the decision point comes after these systems are available. Teams should benchmark complete workflows—data preparation, retrieval, model serving and inter-node communication—then compare cost per completed task, not only advertised chip throughput. Availability, pricing and measured software performance remain open questions.