Advancing model intelligence starts with training, and training starts with data. Before GPU supercomputers can train a model, enormous fleets of CPUs must collect, clean, filter, and format that data. If CPUs fall behind, expensive GPU capacity becomes underutilized and delays the delivery of smarter models to market. As demand grows for models yielding ever more intelligence per task, the need for leadership AI data processing capabilities becomes paramount.
To address that need, Microsoft is excited to announce today the Preview of Azure HDv2 virtual machines (VMs), the fastest and most energy efficient platform ever for AI data processing. HDv2 VMs are built with 6th Gen AMD EPYC™ processors, codenamed “Venice.” Customers can sign up request access to the Preview at aka.ms/HDv2Preview.
“When we prepare large datasets for AI research, more CPU cores let us split the work across more jobs, provided each job has enough memory. That same balance matters when many coding agents build software and run tests at once. We’ve worked with Microsoft on Azure HDv2, and we’re excited to explore what it can bring to our research as these workloads grow.”
-Rafael Dietrich, Head of CPU & Storage Capacity Planning, OpenAI
“As AI scales, data preparation, processing, and orchestration all place enormous demands on CPU compute. With Azure HDv2, AMD and Microsoft co-designed a platform around 6th Gen AMD EPYC processors with Azure’s expertise in cloud infrastructure at scale. Our work together will help customers accelerate turning massive data sets into frontier intelligence.”
-Dan McNamara, Senior Vice President and General Manager of Compute and Enterprise AI, AMD
Enabling Faster Time to Intelligence
HDv2 has been co-designed with leading AI model developers for the highest levels of data processing performance and efficiency. To support highly parallelized applications like Apache Spark, HDv2 features more than 500 physical AMD “Zen 6” CPU cores from the newly introduced EPYC 9006-series of processors. To store vast datasets in memory across a fleet of compute VMs, HDv2 also features 3.8 terabytes of DDR5-8800 memory. To cache these datasets close to compute, HDv2 features 30 TB of PCIe Gen 5 NVMe storage. And to quickly ingress raw data and egress a refined dataset primed for training, HDv2 features 400 Gb Azure Boost Networking. Together, this design enables HDv2 to deliver the highest levels of performance and energy efficiency ever for data processing, even compared against the latest in-market alternatives.
Figure 1: HDv2 VM specifications, with Apache Spark performance, and performance/watt measurements. Preliminary measurements only and subject to change.
Leadership in Compute-bound HPC, and Agentic + RL Task Throughput
Data preparation for AI training is only one of HDv2’s strengths, as it also brings very large performance advantages to other popular HPC and agentic workloads Azure customers rely on for research, development, and mission-critical services.
First, HDv2 delivers large gains for compute-bound HPC applications thanks to its 500 cores capable of single-cycle AVX-512 instruction throughput, which yield more than 2x the peak FP64 performance of any other Cloud dual-socket server. Coupled with 3.8 TB of RAM per VM and 8 GB of RAM per core profile, HDv2 can run the large majority of HPC models from research and industry within a single VM at exceptionally high performance. HPC workloads such as implicit finite element analysis, acoustics, molecular dynamics, and electromagnetics, financial risk simulation, and high-fidelity post-processing will all see significant speedups compared against the latest in-market alternatives.
Below, find a performance comparison across the newly released SPEC CPU 2026 Floating Point benchmark suite evaluating 12 distinct workload domains representing the broader HPC market spanning industrial technical computing, to scientific simulation, and beyond.
Figure 2: HDv2 performance on SPEC CPU 2026 FP Rate, with comparative diverse conventional HPC workload domains. Preliminary measurements only and subject to change.
In addition, HDv2’s more than 500 physical cores make it Azure’s most powerful platform to date for agentic AI and reinforcement learning (RL), where CPUs set the pace for the GPUs they support. Every time an agent runs code or calls a tool, it does so in an isolated sandbox on a CPU, and RL runs hundreds or thousands of these sandboxes at once. When cores run short, the GPUs directing the agents sit idle. A single HDv2 VM can run about twice as many sandboxes in parallel as the next-best in-market alternative with up to 1.7x more agentic task throughput, keeping accelerators productive and agentic workflows responsive to end users. The chart below shows HDv2’s throughput advantage on SPEC CPU 2026 Integer Rate, a proxy for agentic and RL task throughput.
Figure 3: HDv2 performance on SPEC CPU 2026 Integer Rate, with comparative agentic AI and reinforcement learning task-throughput measurements. Preliminary measurements only and subject to change.
Broadly Accessible by Design
HDv2’s advantages for a wide variety of customer workloads means it is important Microsoft make it as accessible as possible. As such, HDv2 will be made available at GA through a broad set of Microsoft services such as Azure Kubernetes Service, Azure Databricks, and Azure CycleCloud, as well as leading partner platforms such as those from Rescale, Synopsys (ANSYS), and others.
Figure 4: Microsoft services and partner platforms for broad access to Azure HDv2 VMs. Availability may vary.
“Azure HDv2 VMs will empower Rescale’s HPC customers with new levels of performance. HDv2’s 500+ cores and nearly 4 TB of memory expand the performance frontier for CAE workflows, from larger implicit FEA models and dense-matrix electromagnetics, to coupled fluid-structure simulations. By putting more compute and memory in a single node, these VMs allow many workloads currently spread across multiple nodes to run on a single node, reducing communication overhead and simplifying execution”
-Adam McKenzie, Chief Technology Officer, Rescale
Sign up for the HDv2 preview
Preview registration is now open for Azure HDv2 VMs. Customers interested in faster AI data preparation, compute-intensive HPC performance, and dense agentic sandbox capacity can sign up today to request access and evaluate HDv2 with their workloads.
Microsoft