{
 "id": "eagle",
 "url": "/supercomputers/eagle",
 "name": "Eagle (Microsoft Azure)",
 "site": "Microsoft Azure",
 "country": "US",
 "compute_atlas_url": "https://computeatlas.net/supercomputers/systems/eagle",
 "sourcing": "Microsoft Azure blog posts (March and August 2023), Microsoft Learn ND H100 v5 documentation, Microsoft's Azure HPC roundup for 2023, NVIDIA's press release and architecture write-ups, and the TOP500 entry, researched 2026-10-10",
 "reading_note": "Unmarked companies come from the operator's or a vendor's own pages (azure.microsoft.com, learn.microsoft.com, techcommunity.microsoft.com, nvidia.com, top500.org); reported companies come from named trade press or analyst reporting whose article body was read in full, not just the headline.",
 "specs": {
  "accelerator": "NVIDIA H100 Tensor Core GPU (SXM, 80 GB HBM3 each); 8 per ND H100 v5 VM, 14,400 in the TOP500 run",
  "host_cpu": "Intel Xeon Platinum 8480C (4th Gen Xeon Scalable, Sapphire Rapids), listed by TOP500 as 48 cores at 2 GHz; 96 physical host cores per ND H100 v5 VM",
  "integrator": "Microsoft (TOP500 lists the manufacturer as Microsoft Azure; server OEM or ODM not disclosed)",
  "nodes": "About 1,800 ND H100 v5 nodes (14,400 H100 GPUs at 8 per VM, derived); TOP500 lists 2,073,600 cores (1,123,200 on the November 2023 list)",
  "memory": "Per ND H100 v5 VM: 1,900 GiB host DDR5 (16 channels of 4800 MHz DDR5 DIMMs) and 8 x 80 GB HBM3 (640 GB GPU memory); system total not published",
  "interconnect": "NVIDIA Quantum-2 InfiniBand NDR, one 400 Gb/s ConnectX-7 adapter per GPU (3.2 Tb/s per VM) in a non-blocking fat-tree; NVSwitch and NVLink 4.0 inside each VM (3.6 TB/s bisectional bandwidth between the 8 GPUs)",
  "storage": "Per ND H100 v5 VM: 1 x 1,024 GiB temp disk, 8 local NVMe disks (28 TiB listed) and up to 32 remote managed disks; the shared storage behind Eagle is not published",
  "rmax_pflops": "561.20 PFlop/s (TOP500 June 2026 list)",
  "rpeak_pflops": "846.84 PFlop/s (TOP500 June 2026 list)",
  "status_note": "Microsoft describes Eagle as Azure's own cloud supercomputer, built from ND H100 v5 virtual machines (preview March 2023, general availability 7 August 2023), and says the same capability is available to anyone with a credit card on Azure; TOP500 lists it as 'Microsoft NDv5' at site Microsoft Azure (installation year 2023, first listed November 2023, still listed June 2026). Microsoft's CTO said the 14,400-GPU TOP500 run was a fraction of the system's ultimate scale. Not the NREL Eagle."
 },
 "not_confirmed": "Microsoft does not say who builds the ND H100 v5 servers (no OEM or ODM is named in any source read), which datacenter or datacenters host the Eagle cluster, what cools it, how much power the Linpack run drew, or what storage sits behind it; the TOP500 entry lists no power figure. The DRAM, NVMe and SmartNIC vendors, the InfiniBand optics and cable vendors, and the makers of the HGX-class GPU baseboard are also unnamed, so none of those get a line. Microsoft describes the TOP500 run as a fraction of the system's ultimate scale and says its AI capacity has since grown by orders of magnitude, so 14,400 GPUs is the size of the benchmarked configuration, not a statement of the whole fleet. The TOP500 entry shows the same Rmax on every list from November 2023 to June 2026, with the core count revised from 1,123,200 to 2,073,600 after the first list.",
 "lines": [
  {
   "n": 1,
   "line": "System design, operation and the ND H100 v5 node (8x H100 GPUs, 96 Xeon cores, 1,900 GiB DDR5, 8x 400 Gb/s InfiniBand adapters per VM; about 1,800 nodes in the TOP500 run)",
   "parts": [
    {
     "id": "9.02",
     "name": "Hyperscalers, neoclouds & model labs",
     "url": "/parts/building-shell-site-and-construction/hyperscalers-neoclouds-and-model-labs-the-tenants-and-owners"
    }
   ],
   "companies": [
    {
     "id": "microsoft",
     "name": "Microsoft",
     "url": "/companies/microsoft",
     "reported": false,
     "note": "designs, builds and operates the Azure ND H100 v5 cluster and submitted it to TOP500 as \"Microsoft NDv5\"; TOP500 lists the manufacturer as Microsoft Azure"
    }
   ],
   "unresolved": [
    "OEM or ODM that builds the ND H100 v5 servers and racks (no source read names one)",
    "GPU baseboard maker for the 8-GPU board (Microsoft's pages do not say HGX or name a supplier)",
    "datacenter site or sites that host the Eagle cluster (not disclosed)"
   ],
   "sources": [
    "https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/ndh100v5-series",
    "https://top500.org/system/180236/",
    "https://www.top500.org/lists/top500/2023/11/highs/",
    "https://techcommunity.microsoft.com/blog/azurehighperformancecomputingblog/annual-roundup-of-ai-infrastructure-breakthroughs-for-2023/4097737",
    "https://thenewstack.io/sc500-microsoft-now-has-the-third-fastest-computer-in-the-world/"
   ]
  },
  {
   "n": 2,
   "line": "NVIDIA H100 Tensor Core GPU (SXM, 80 GB, 8 per VM; 14,400 in the TOP500 run)",
   "parts": [
    {
     "id": "1.01",
     "name": "Merchant GPUs / accelerators",
     "url": "/parts/compute-silicon-and-packaging/merchant-gpus-accelerators"
    }
   ],
   "companies": [
    {
     "id": "nvidia",
     "name": "Nvidia",
     "url": "/companies/nvidia",
     "reported": false,
     "note": "H100 GPUs, NVSwitch and NVLink 4.0 baseboard fabric"
    }
   ],
   "unresolved": [],
   "sources": [
    "https://techcommunity.microsoft.com/blog/azurehighperformancecomputingblog/annual-roundup-of-ai-infrastructure-breakthroughs-for-2023/4097737",
    "https://blogs.nvidia.com/blog/efficient-ai-supercomputers-sc23/"
   ]
  },
  {
   "n": 3,
   "line": "Hopper GPU die fabrication (TSMC 4N process)",
   "parts": [
    {
     "id": "1.09",
     "name": "Leading-edge foundry",
     "url": "/parts/compute-silicon-and-packaging/leading-edge-foundry"
    }
   ],
   "companies": [
    {
     "id": "tsmc",
     "name": "TSMC",
     "url": "/companies/tsmc",
     "reported": false,
     "note": "fabricates the GH100 die in the H100 on its 4N process, per Nvidia's Hopper architecture write-up; Microsoft's pages do not name the foundry"
    }
   ],
   "unresolved": [
    "advanced-packaging flow and substrate suppliers for the H100 SXM module (none named in any source read)"
   ],
   "sources": [
    "https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/"
   ]
  },
  {
   "n": 4,
   "line": "GPU memory (80 GB HBM3 per H100, 640 GB per VM)",
   "parts": [
    {
     "id": "2.01",
     "name": "HBM",
     "url": "/parts/memory-and-storage/hbm-hbm3e-hbm4"
    }
   ],
   "companies": [
    {
     "id": "sk-hynix",
     "name": "SK Hynix",
     "url": "/companies/sk-hynix",
     "reported": true,
     "note": "reported as the primary HBM3 supplier for Nvidia's H100; no source read confirms it for the GPUs in Azure's ND H100 v5 servers"
    }
   ],
   "unresolved": [
    "which HBM3 vendors supplied the stacks on Azure's H100 GPUs (Microsoft and Nvidia do not say; TrendForce also notes Samsung entered Nvidia's HBM3 supply in late 2023)"
   ],
   "sources": [
    "https://www.trendforce.com/presscenter/news/20240313-12075.html",
    "https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/ndh100v5-series"
   ]
  },
  {
   "n": 5,
   "line": "Host CPU (Intel Xeon Platinum 8480C, 4th Gen Xeon Scalable / Sapphire Rapids; 96 physical cores per VM)",
   "parts": [
    {
     "id": "1.05",
     "name": "Server CPUs",
     "url": "/parts/compute-silicon-and-packaging/server-cpus"
    }
   ],
   "companies": [
    {
     "id": "intel",
     "name": "Intel",
     "url": "/companies/intel",
     "reported": false,
     "note": "host processors for every ND H100 v5 node"
    }
   ],
   "unresolved": [
    "DDR5 DIMM vendor for the 1,900 GiB of host memory per VM (not named)"
   ],
   "sources": [
    "https://top500.org/system/180236/",
    "https://azure.microsoft.com/en-us/blog/scale-generative-ai-with-new-azure-ai-infrastructure-advancements-and-availability/",
    "https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/ndh100v5-series"
   ]
  },
  {
   "n": 6,
   "line": "InfiniBand adapters (ConnectX-7, 400 Gb/s NDR, one per GPU, 8 per VM)",
   "parts": [
    {
     "id": "1.06",
     "name": "NICs / SuperNICs",
     "url": "/parts/compute-silicon-and-packaging/nics-supernics"
    }
   ],
   "companies": [
    {
     "id": "nvidia",
     "name": "Nvidia",
     "url": "/companies/nvidia",
     "reported": false,
     "note": "Quantum-2 ConnectX-7 InfiniBand adapters"
    }
   ],
   "unresolved": [],
   "sources": [
    "https://azure.microsoft.com/en-us/blog/scale-generative-ai-with-new-azure-ai-infrastructure-advancements-and-availability/",
    "https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/ndh100v5-series"
   ]
  },
  {
   "n": 7,
   "line": "In-node GPU fabric (NVSwitch and NVLink 4.0, 3.6 TB/s bisectional bandwidth between the 8 GPUs of each VM)",
   "parts": [
    {
     "id": "3.01",
     "name": "Scale-up fabric",
     "url": "/parts/networking-and-interconnect/scale-up-fabric-gpu-to-gpu-inside-the-rack-pod"
    }
   ],
   "companies": [
    {
     "id": "nvidia",
     "name": "Nvidia",
     "url": "/companies/nvidia",
     "reported": false,
     "note": "NVSwitch and NVLink 4.0, part of the H100 SXM platform"
    }
   ],
   "unresolved": [],
   "sources": [
    "https://azure.microsoft.com/en-us/blog/azure-previews-powerful-and-scalable-virtual-machine-to-help-customers-accelerate-ai/"
   ]
  },
  {
   "n": 8,
   "line": "InfiniBand fabric (NVIDIA Quantum-2 NDR switching in a non-blocking fat-tree; 3.2 Tb/s per VM)",
   "parts": [
    {
     "id": "3.02",
     "name": "Scale-out switch ASICs",
     "url": "/parts/networking-and-interconnect/scale-out-switch-asics"
    },
    {
     "id": "3.03",
     "name": "Switch systems & white-box ODMs",
     "url": "/parts/networking-and-interconnect/switch-systems-and-white-box-odms"
    }
   ],
   "companies": [
    {
     "id": "nvidia",
     "name": "Nvidia",
     "url": "/companies/nvidia",
     "reported": false,
     "note": "Quantum-2 InfiniBand switching platform; Microsoft names Quantum-2 InfiniBand networking but does not itemise switch models or counts"
    }
   ],
   "unresolved": [
    "switch counts and models, fabric size and layers (not published)",
    "optical transceiver, DAC and fibre cable vendors (not named; Microsoft's CTO says the InfiniBand cabling would wrap around the Earth at least five times, which is a fleet-wide claim)"
   ],
   "sources": [
    "https://azure.microsoft.com/en-us/blog/azure-previews-powerful-and-scalable-virtual-machine-to-help-customers-accelerate-ai/",
    "https://nvidianews.nvidia.com/news/nvidia-microsoft-accelerate-cloud-enterprise-ai",
    "https://top500.org/system/180236/"
   ]
  }
 ]
}