(858) 613-1818Quote list 0
Available on request

8 node HGX H200 AI pod, 64 GPUs

Eight HGX H200 servers with 64 Hopper GPUs and 1.1 TB of HBM3e per node on a 400 or 800 Gb/s fabric, with control plane and flash storage, on request.

GPU: NVIDIA HGX H200

Illustration of one rack with eight GPU nodes and a top of rack switch
GPU nodes
8 x 8 GPU HGX H200 servers, 64 x H200 SXM, 8.8 TB HBM3e (8 x 1.1 TB per node)
Fabric
64 x 400 Gb/s ConnectX-7 NDR InfiniBand or 400 GbE, or 64 x 800 Gb/s ConnectX-8, non blocking
CPU and control
2 x CPU nodes for data services, 6 x control plane nodes, AMD EPYC 9005 or Intel Xeon 6
Storage
64 TB usable all flash minimum, 1 TB per GPU, 2 x NVMe storage nodes
Racks
3 to 4 racks at about 10 kW per GPU node, about 95 kW total
Delivery
Supplied on request as Hopper allocation allows, racked, cabled and fabric validated

Overview

The 8 node HGX H200 pod applies our standard pod design to the Hopper platform for teams that are expanding an existing H100 or H200 cluster or whose software stack is validated on Hopper and will not move soon. Eight HGX H200 servers with 64 GPUs and 1.1 TB of HBM3e per node sit on a non blocking fabric with the CPU, control plane and storage nodes a production cluster needs.

We list it as available on request because Hopper allocation now follows the Blackwell ramp and configurations vary by batch. The lower planning power of about 10 kW per node, against 14 kW for Blackwell nodes, makes it the easier pod to place in a hall with limited rack density, and 400 Gb/s ConnectX-7 adapters let it join an NDR InfiniBand or 400 GbE fabric that is already in place.

Sizing follows the same rules as every cluster we deliver: six CPU threads and 24 GB of host memory per GPU, six control plane nodes and one terabyte of usable flash per GPU. We rack, cable and validate before shipment, provide the elevation and power schedule for your data center or colocation facility, and offer on site installation, with the shipment timed to your ready date.

Use cases

  • Expanding Hopper clusters
  • Inference on validated Hopper stacks
  • Research and university clusters
  • Halls with limited power per rack

Specifications

Compute
GPU nodes8 x 8 GPU HGX H200 AI server, air cooled, 8U each
GPUs64 x NVIDIA H200 SXM, 141 GB HBM3e each, 8.8 TB total (8 x 1.1 TB per node)
GPU interconnectFourth generation NVLink, 900 GB/s per GPU, NVSwitch within each node
CPU nodes2 x 1U dual socket EPYC 9005 compute servers (384 threads and 1.5 TB memory each, meeting 6 threads and 24 GB per GPU)
Control plane6 x 1U single socket EPYC 9005 control plane servers
Intel optionGPU node hosts, CPU nodes and control plane nodes can be specified on Intel Xeon 6 instead of AMD EPYC 9005
Fabric
East west64 x 400 Gb/s ConnectX-7 on NDR InfiniBand or 400 GbE switches, or 64 x 800 Gb/s ConnectX-8 on Quantum-X800 or Spectrum-X class switches, non blocking
North south16 x 400 Gb/s BlueField-3 or ConnectX-7 ports into a 400 GbE leaf pair
Management1 GbE out of band network for every BMC and switch
Cabling400G QSFP112 or 800G OSFP transceivers and cables, lengths per rack elevation
Storage
Capacity64 TB usable minimum, 1 TB per GPU, expandable
Nodes2 x 2U 24 bay all NVMe storage servers with 2 x 400 Gb/s each
SoftwareParallel file system or NVMe over Fabrics, specified with the order
Facility
Racks3 to 4 x 42U or 48U racks, 3 GPU nodes per rack at about 30 kW, plus 1 rack for storage, CPU, control plane and switches
Planning powerAbout 95 kW total: 8 x 10 kW GPU nodes, 2 x 1 kW CPU, 6 x 0.5 kW control, 2 x 1.3 kW storage, switches and overhead
CoolingAir cooled, hot aisle containment recommended
Platform options
Standard8 x HGX H200 nodes, 64 GPUs, air cooled, on request
Blackwell pathHGX B200 or HGX B300 nodes join the same fabric design; see the 8 node HGX B200 and HGX B300 AI pods
ChoiceDecided with the fabric design and the delivery date

Figures describe the system class. Exact configuration, weights and power draw are confirmed on the quotation.

What's included

  • Rack elevation, cabling plan and power schedule
  • Nodes, switches, cables and optics as one bill of materials
  • Burn in of every node and fabric validation across the cluster, acceptance report
  • NVIDIA end user registration handled with the order
  • Deployment support and optional on site installation
  • 3 year warranty, options to 5 years with advance replacement

Compliance

Compare it with the next class up or down

The nearest classes we quote are the 8 GPU HGX H200 AI server and the 8 node HGX B200 AI pod, 64 GPUs. If neither is quite right, tell us the workload and we specify the system between them.

Talk to an engineerRequest a quotation