Microsoft announced the expansion of Azure’s AI and high-performance computing infrastructure through the integration of AMD’s latest Helios AI platform and next-generation data center EPYC processors. These technologies will power three upcoming Azure offerings: HDv2 for data processing, HXv2 for electronic design automation, and ND MI455X v7 for AI inference workloads.
The move comes as AI workloads expand at a pace that no single infrastructure can meet, with the number of models increasing, new agent-based workloads emerging, and demand for computing capacity rising. Microsoft says Azure is moving toward an open platform with diverse components, combining partner technologies such as AMD with chips and systems developed internally by Microsoft, with the goal of giving customers broader options for balancing performance, cost, and energy efficiency.
HDv2 for Processing AI System Data
Azure HDv2 virtual machines are designed to address bottlenecks associated with CPU-based infrastructure in AI systems. Accelerators require dense, energy-efficient computing capacity to process data, coordinate workloads, and maintain the continuity of processing pipelines, while training and agent tasks need sufficient capacity to feed models and execute tasks.
HDv2 was developed in collaboration with AMD to support workloads such as data preparation, search, reinforcement learning, and large-scale agent orchestration. The virtual machines include approximately 500 physical cores from sixth-generation AMD EPYC processors, 4 terabytes of RAM, and 32 terabytes of local NVMe storage, along with 400-gigabit Azure Boost network connectivity.
HXv2 for Chip Design and Technical Computing
HXv2 builds on the HX series that Azure launched in partnership with AMD in 2023, which uses AMD’s 3D V-Cache technology and has seen adoption among chip design companies. HXv2 continues to support RTL simulation workloads, with improvements in single-thread performance and memory.
HXv2 includes up to 176 cores from sixth-generation AMD EPYC processors clocked above 5 GHz, along with a 50% increase in addressable cache per core. Virtual machine sizes with nearly 2 or 4 terabytes of memory are also available, helping customers align memory resources with their workload requirements.
HXv2 is not limited to chip design; it also targets scientific simulation, engineering analysis, and distributed-memory applications. The virtual machine provides 800-gigabit InfiniBand connectivity, supporting large-scale MPI-based simulations and enhancing its suitability for a variety of high-performance computing workloads.
Mark Papermaster, executive vice president and chief technology officer at AMD, said that engineering teams face growing requirements in simulation, chip design, and scientific computing, adding that Azure HXv2 is designed to deliver greater performance and scalability. Shankar Krishnamurthy, chief product development officer at Synopsys, also noted that the company’s collaboration with Microsoft on the Azure HX series helps its customers scale electronic design automation workloads through cloud computing.
ND MI455X v7 for AI Inference
The ND MI455X v7 virtual machine is designed for inference, search, and agent-based workloads that underpin modern AI services. It relies on AMD’s rack-scale Helios solution, with the aim of expanding Azure’s options for large-scale inference and providing performance and efficiency suited to demanding workloads.
According to Microsoft, the three offerings give customers greater flexibility to select computing resources according to the required workflow, from inference, through AI data systems, to chip design. The announcement did not specify when these offerings would become available or provide pricing details.