TL;DR<\/p>\n
Azure Expansion: Microsoft plans to deploy AMD\u2019s Helios rack-scale AI system on Azure and broaden AMD\u2019s infrastructure role. Helios Stack: Helios combines MI455X accelerators, EPYC Venice processors, Pensando networking and Radeon Open Compute software in one coordinated rack. Azure Services: Azure also plans HDv2 for data pipelines, HXv2 for chip design and ND MI455X v7 for AI inference. Deployment Limits: Financial terms, deployment capacity, scale-up networking and independent benchmarks remain unavailable ahead of planned second-half 2026 shipments. <\/p>\n
Microsoft and AMD have expanded their Azure partnership across accelerators, processors, networking and software.<\/p>\n
Microsoft plans to deploy AMD\u2019S Helios rack-scale AI supercomputing platform on Azure and expand its Azure AI and HPC infrastructure, broadening AMD\u2019s role across accelerators, processors, networking and software.<\/p>\n
For Azure customers, the expansion creates another integrated option for running trained AI models and related services, although Microsoft and AMD did not disclose financial terms or deployment capacity. Futurum Group\u2019s estimate would put Nvidia\u2019s GPU market share at more than 95%<\/a> and AMD\u2019s at about 4.5%. One large cloud customer does not erase that competitive gap, but it gives AMD a prominent environment in which to test a complete system.<\/p>\n What Microsoft Is Adding to Azure <\/p>\n Helios integrates MI455X, Venice, Pensando and ROCm<\/a> in one rack-scale platform. A rack-scale system coordinates processors, accelerators, networking and software as one unit instead of treating each component as a separate server purchase.<\/p>\n MI455X accelerators provide the AI compute, EPYC Venice CPUs handle general processing, Pensando manages network traffic, and Radeon Open Compute (ROCm)<\/a> supplies AMD\u2019s accelerator software layer. Memory, processor scheduling, network traffic and software must work together before Azure can offer dependable model inference.<\/p>\n Helios targets 260 TB\/s and 43 TB\/s<\/a> for data movement within racks and between racks, respectively. The first figure covers communication among accelerators in one system, while the second covers traffic across multiple racks.\u00a0<\/p>\n Azure also plans three AMD-powered offerings<\/a> with different jobs. HDv2 virtual machines (VM) target agentic AI and data pipelines, HXv2 VMs target chip design, and ND MI455X v7 will use Helios to execute trained AI models. Azure Foundry Managed Compute adds a managed route to AMD-powered infrastructure, subject to Microsoft\u2019s rollout schedule and AMD\u2019s shipment plan.<\/p>\n HDv2 includes nearly 500 physical EPYC cores, 4 TB of memory, 32 TB of local NVMe storage and 400 Gb Azure Boost networking. HXv2 includes 176 EPYC cores running above 5 GHz, 50% more addressable cache per core than HX, nearly 2 TB or 4 TB of memory, and 800 Gb InfiniBand. Those configurations distinguish CPU-heavy services from the Helios inference system and give chip-design workloads a separate high-frequency option.<\/p>\n AMD\u2019s Pensando data processing units (DPUs)<\/a> offload networking and storage work from central processing units. Moving connection handling away from the main processors leaves more CPU capacity for customer workloads. Microsoft already uses Azure Boost\u2019s DPU architecture<\/a> since 2024; the planned AMD integration extends that established offload mechanism instead of adding another accelerator alone.<\/p>\n The Competitive Test Still Lacks Deployment Proof <\/p>\n Microsoft also expands Azure AI infrastructure work with Nvidia, making the AMD plan an addition to a multi-supplier strategy rather than a replacement. Nvidia\u2019s Compute Unified Device Architecture (CUDA) gives developers an established software ecosystem around its accelerators. Counterpoint Research analyst Neil Shah attributes Nvidia\u2019s larger and materially more advanced ecosystem to CUDA<\/a>.<\/p>\n Scarce AI capacity can generate hardware demand without demonstrating a performance advantage. Futurum Group analyst and CEO Daniel Newman sees AMD\u2019s gains as potentially reflecting either technological superiority or capacity constraints that ensure available hardware finds customers.<\/p>\n Helios enters a rack-system contest in which AMD must establish accelerator capability and a dependable software and networking package. Microsoft\u2019s purchase gives AMD a prominent customer test, not proof that Helios has closed the software or performance gap.<\/p>\n Microsoft has not publicly specified the scale-up network for its initial Helios racks. The interconnect controls how accelerators exchange data inside a rack, making its design material to sustained system performance.<\/p>\n Once AMD begins Helios shipments in the second half of 2026, Microsoft still needs to define its commitment by disclosing the number of systems assigned to Azure.<\/p>\n","protected":false},"excerpt":{"rendered":"TL;DR Azure Expansion: Microsoft plans to deploy AMD\u2019s Helios rack-scale AI system on Azure and broaden AMD\u2019s infrastructure…\n","protected":false},"author":2,"featured_media":116265,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[11],"tags":[4518,9680,14512,406,8158,407,13974,14952,3125,57815,420,7829,2304,425,2193,39,57649,412,59155,320,7828,4528,15967,3444,58,33981,59156,3318,59157,2011,59158],"class_list":["post-116264","post","type-post","status-publish","format-standard","has-post-thumbnail","category-microsoft","tag-ai-chips","tag-ai-competition","tag-ai-compute","tag-ai-hardware","tag-ai-inference","tag-ai-infrastructure","tag-ai-partnerships","tag-ai-services","tag-amd","tag-amd-epyc","tag-azure","tag-azure-ai","tag-cloud-infrastructure","tag-cloud-computing","tag-cpus","tag-data-centers","tag-epyc-venice","tag-gpus","tag-mi455x","tag-microsoft","tag-microsoft-ai","tag-microsoft-azure","tag-microsoft-cloud","tag-networking","tag-nvidia","tag-nvidia-cuda","tag-pensando-data-processing-unit-dpu","tag-processors","tag-radeon-open-compute-rocm","tag-semiconductors","tag-virtual-machines-vms"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/116264","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=116264"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/116264\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/116265"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=116264"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=116264"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=116264"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}