Microsoft picks AMD Helios racks for Azure AI

Microsoft and AMD announced a large-scale Azure deployment built around AMD’s “Helios” rack design, combining “Altair” MI455X GPUs, “Venice” Epyc 9006 CPUs, Pensando DPUs and the ROCm software stack. The systems are targeted specifically at inference for frontier models on the Azure cloud, adding another major cloud deployment for AMD’s AI hardware platform.

The double-wide Helios rack has 4,600 Zen 6 CPU cores across 18 compute trays, with one CPU and four GPUs each. Each Venice processor has 256 cores, and a rack includes 18,000 GPU compute units across 72 GPUs. The GPUs deliver 2.9 exaflops at FP4 precision, with a combined 31 TB of HBM 4 stacked memory and 43 TB/sec of aggregate bandwidth through Pensando DPUs programmable in the P4 language.

Microsoft also committed to clusters of Venice Epyc CPUs for two Azure instance families. Azure HDv2 instances will target agentic AI workloads and data pipeline processing, while HXv2 instances will focus on electronic design automation for chips. Microsoft is also working with AMD to move Azure Boost acceleration software to Pensando DPUs, shifting networking and storage virtualization functions onto AMD’s programmable data processing hardware.