Microsoft adds AMD Helios and EPYC 6 to Azure AI infrastructure

Microsoft diversifies Azure compute with AMD's latest silicon

Microsoft is integrating AMD's newest AI and high-performance computing technologies into Azure, introducing three virtual machine families designed for distinct workload categories that have emerged as AI systems scale beyond training alone.

The expansion, announced by Scott Guthrie, Microsoft's Executive Vice President of Cloud + AI, brings AMD's Helios AI platform and 6th Generation EPYC datacenter processors to Azure through HDv2, HXv2, and ND MI455X v7 virtual machines. Each targets a specific bottleneck in modern AI infrastructure: data pipeline processing, electronic design automation, and production-scale inference.

Why it matters

As AI workloads fragment into specialized tasks—agent coordination, reinforcement learning, chip simulation, reasoning engines—cloud providers face pressure to offer more than generic GPU instances. This move signals that hyperscalers are building heterogeneous fleets where CPU architecture, memory configuration, and interconnect topology matter as much as accelerator choice. For enterprises running complex AI systems, it means more granular control over performance-cost tradeoffs across the full stack.

Three workload-specific VM families

Azure HDv2 virtual machines address a constraint that has limited agentic AI adoption: insufficient CPU capacity to feed data pipelines and coordinate multi-step workflows. These VMs pack nearly 500 physical AMD EPYC cores, 4 terabytes of RAM, 32 terabytes of local NVMe storage, and 400 Gb Azure Boost networking. Microsoft positions them for data preparation, search indexing, reinforcement learning environments, and agent orchestration at scale.

HXv2 VMs target silicon design firms and technical computing users. Building on the HX series launched with AMD in 2023, HXv2 employs AMD's 3D V-cache technology optimized for RTL simulation workloads. The new generation features 176 6th Gen EPYC cores running above 5 GHz, 50 percent more addressable cache per core, and configurations with up to 4 terabytes of RAM. The addition of 800 Gb InfiniBand support extends HXv2's utility to large-scale MPI-based scientific simulations beyond chip design.

AMD itself uses Azure HX-series infrastructure to design future EPYC CPUs and Instinct GPUs, according to Mark Papermaster, AMD's Executive Vice President and CTO. Synopsys, whose AI-powered EDA tools run on the platform, noted that the collaboration enables customers to extend workloads beyond traditional infrastructure constraints while meeting aggressive development schedules.

ND MI455X v7 VMs are built for AI inference workloads—specifically reasoning, search, and agentic services. Powered by AMD's Helios rackscale solution, these instances expand Azure's options for deploying large-scale inference infrastructure with what Microsoft describes as strong performance and efficiency characteristics.

Heterogeneous infrastructure as strategy

Microsoft frames the AMD integration as part of a broader infrastructure philosophy that combines third-party silicon, purpose-built accelerators, and custom systems. Guthrie emphasized that customer choice remains a core design principle, with different compute architectures suited to different AI workflow stages.

The announcement reflects a market reality: no single chip architecture efficiently handles the full spectrum of AI workloads. Training, fine-tuning, data preprocessing, inference, and agent coordination each have distinct compute profiles. Cloud providers that offer granular infrastructure options can capture more of the AI stack—and more margin—than those selling undifferentiated GPU hours.

Details were first reported by Microsoft on the Official Microsoft Blog.