Visit Azure and AMD at Microsoft Ignite and SC26

AI workloads place different demands on infrastructure. Azure is continuing to partner closely with silicon partners including AMD to expand infrastructure choice across general purpose compute and AI accelerators. That choice includes three purpose built virtual machines (VMs): Azure HDv2-series virtual machines for the data preparation behind AI training, AMD Instinct™ MI300X accelerators for AI Inference and agents, generally available as managed endpoints in Microsoft Foundry or as Azure ND MI300X v5 VMs, and AMD Instinct™ MI455X accelerators in Azure for next generation AI infrastructure coming to Azure.

In July, we shared how Azure is expanding its AI and HPC infrastructure with AMD. Here’s where to learn more at Microsoft Ignite and SC26.

Azure HDv2: accelerating AI data processing, HPC, and agentic workloads 

Accelerators are only as productive as the data feeding them, and CPUs play an integral role in preparing that data. Microsoft has opened preview registration for Azure HDv2-series virtual machines, making Azure the first cloud to offer access to 6th Gen AMD EPYC™ (“Venice”) processors. HDv2 is built for AI data systems, with over 500 physical CPU cores, 3.8 TB of memory, 32 TB of local NVMe storage and 400 Gb Azure Boost networking.

In benchmark testing on a preproduction AMD system with two “Venice” processors, an AI data-preparation pipeline prepared a 1.3 TB dataset for AI training in about 37 minutes¹. HDv2 is projected to cut runtime by 49% compared with the next-best tested configuration (about 73 minutes), helping teams start training sooner and make better use of accelerator capacity. With over 500 cores, HDv2 can also speed compute-intensive HPC workloads such as engineering simulation and weather modeling. Customers can sign up to request access at aka.ms/HDv2Preview.

AMD Instinct™ MI300X, on Azure today: a managed endpoint or your own VM

AMD Instinct™ MI300X is generally available on Azure today. Through Microsoft Foundry Managed Compute, teams can deploy open models from the Foundry model catalog onto dedicated MI300X capacity, via a private endpoint. Microsoft manages the software needed to run each model, so teams do not need to set up or maintain it themselves. Capacity is billed by the GPU-hour, and billing stops when the instances are spun down deployment is deleted.

In a third-party evaluation by Signal65, AMD Instinct™ MI300X on Microsoft Foundry Managed Compute deployed live endpoints in 14-24 minutes with no serving stack tuning across the leading family of models including gpt-oss, gemma and GLM. MI300X delivered strong results across performance, cost per session, and agent capacity for coding, document analysis and chat workloads².

Teams that want full control of the software stack, or that are fine-tuning and training their own models can use Azure ND MI300X v5 virtual machines, with eight AMD Instinct™ MI300X accelerators and 1.5 TB of HBM per VM. Each MI300X provides 192 GB of HBM3 memory, allowing large models to run on fewer accelerators.

AMD Instinct™ MI455X accelerators in Azure: engineered as a complete system

That same focus on purpose-built infrastructure extends to AI accelerators. Azure and AMD are working together on AI infrastructure built around AMD Instinct™ MI455X accelerators and AMD’s Helios™ reference design.

Running AI at scale requires more than AI accelerator performance. Consistent with how we approach AI infrastructure, Azure and AMD work together to design this complete system, so it fits into our Azure datacenter without retrofitting, with power, cooling and networking designed and validated together as one system. Microsoft’s system-level engineering across custom and partner silicon helps us bring advanced AI infrastructure from production to deployment faster, meaning we deliver more capacity to meet Azure and AI demand.

Join us at Microsoft Ignite and SC26 to learn more about how Azure and AMD are expanding infrastructure choice.

1 Projected HDv2 performance is based on benchmark results from a preproduction AMD system with two “Venice” processors, using recommended software and settings. Results may vary depending on workload, software, configuration and the final Azure environment.

2 Study commissioned by Microsoft. Signal65, Tables 3, 5–8, 8–9. GPU-only economics at sustained tested load and $7.91/GPU-hour. Throughput is aggregate. Gemma chat economics calculated from 1,065 output tok/s and 490 output tokens/request. 1K-input/490-output test. Different workloads; no quality ranking. Use cases are illustrative.