8 node HGX B200 AI pod, 64 GPUs
Eight HGX B200 servers with 64 GPUs and 1.4 TB of HBM3e per node on one 800 or 400 Gb/s fabric, with control plane and all flash storage, build to order.
GPU: NVIDIA HGX B200
- GPU nodes
- 8 x 8 GPU HGX B200 servers, 64 x B200, 11.2 TB HBM3e (8 x 1.4 TB per node)
- Fabric
- 64 x 800 Gb/s ConnectX-8 ports, or 64 x 400 Gb/s ConnectX-7 for existing NDR fabrics, 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 14 kW per GPU node, about 125 kW total
- Delivery
- Build to order without allocation, racked, cabled and fabric validated, optional on site installation
Overview
The 8 node HGX B200 pod is the same deployable cluster design as our HGX B300 pod, built on the Blackwell platform that ships without allocation. Eight HGX B200 servers with 64 GPUs and 1.4 TB of HBM3e per node sit on a non blocking fabric, with the CPU, control plane and storage nodes that a production cluster needs, specified against the same sizing rules we use for every cluster.
It is the pod to choose when delivery timing matters more than the extra memory and dense FP4 throughput of B300, when a cluster is already tuned for Blackwell, or when the pod will join an existing NDR InfiniBand or 400 GbE fabric and 400 Gb/s ConnectX-7 adapters are the better fit. The sizing is unchanged: 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 the fabric before shipment, provide the elevation and power schedule for your data center or colocation facility, and offer on site installation. Plan for three to four racks and about 125 kW. B300 nodes can be added to the same fabric later, so the pod you buy today grows into the next generation.
Use cases
- Training models up to the 100B parameter class
- Production inference clusters on Blackwell
- Extending existing Blackwell or NDR InfiniBand deployments
- Enterprise AI platforms with a fixed delivery date
Specifications
| GPU nodes | 8 x 8 GPU HGX B200 AI server, air cooled, 8U each |
|---|---|
| GPUs | 64 x NVIDIA B200, 180 GB HBM3e each, 11.2 TB total (8 x 1.4 TB per node) |
| FP4 performance | 144 PFLOPS sparse per node, 1,152 PFLOPS sparse across the pod |
| CPU nodes | 2 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 plane | 6 x 1U single socket EPYC 9005 control plane servers |
| Intel option | GPU node hosts, CPU nodes and control plane nodes can be specified on Intel Xeon 6 instead of AMD EPYC 9005 |
| East west | 64 x 800 Gb/s on Quantum-X800 or Spectrum-X class switches, or 64 x 400 Gb/s NDR InfiniBand or 400 GbE where the pod joins an existing fabric, non blocking |
|---|---|
| North south | 16 x 400 Gb/s BlueField-3 ports into a 400 GbE leaf pair |
| Management | 1 GbE out of band network for every BMC and switch |
| Cabling | 800G OSFP or 400G QSFP112 transceivers and cables, lengths per rack elevation |
| Capacity | 64 TB usable minimum, 1 TB per GPU, expandable |
|---|---|
| Nodes | 2 x 2U 24 bay all NVMe storage servers with 2 x 400 Gb/s each |
| Software | Parallel file system or NVMe over Fabrics, specified with the order |
| Racks | 3 to 4 x 42U or 48U racks, 3 GPU nodes per rack at about 42 kW, plus 1 rack for storage, CPU, control plane and switches |
|---|---|
| Planning power | About 125 kW total: 8 x 14 kW GPU nodes, 2 x 1 kW CPU, 6 x 0.5 kW control, 2 x 1.3 kW storage, switches and overhead |
| Cooling | Air cooled, hot aisle containment recommended |
| Standard | 8 x HGX B200 nodes, 64 GPUs, air cooled, build to order |
|---|---|
| Upgrade path | HGX B300 nodes join the same fabric design when allocation allows; see the 8 node HGX B300 AI pod |
| Choice | Decided 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
Related
Related systems
8 GPU HGX B200 AI server
- GPUs
- 8 x NVIDIA B200, 180 GB HBM3e each
- CPU
- 2 x AMD EPYC 9005 or 2 x Intel Xeon 6
- Memory
- Up to 3 TB DDR5 RDIMM
- Chassis
- 8U, about 14 kW, air cooled
8 node HGX B300 AI pod, 64 GPUs
Allocation required- GPU nodes
- 8 x 8 GPU HGX B300 servers, 64 x B300
- Fabric
- 64 x 800 Gb/s ports, non blocking
- CPU and control
- 2 x CPU nodes for data services
- Chassis
- Multi rack, about 125 kW, air cooled
8 node HGX H200 AI pod, 64 GPUs
Available on request- GPU nodes
- 8 x 8 GPU HGX H200 servers, 64 x H200 SXM
- Fabric
- 64 x 400 Gb/s ConnectX-7 NDR InfiniBand
- CPU and control
- 2 x CPU nodes for data services
- Chassis
- Multi rack, about 95 kW, air cooled
ConnectX-7 400 Gb/s adapter
- Speed
- 400 Gb/s single port, or 2 x 200 Gb/s
- Host interface
- PCIe Gen5 x16
- Offloads
- RDMA and RoCE, GPUDirect RDMA
- Form factor
- PCIe HHHL card with OSFP or QSFP112 port
Related guides
Sizing CPU, memory and storage around GPUs
Six rules of thumb that keep the GPUs fed, with the arithmetic for an 8 GPU node, a PCIe inference server and a 16 node training pod.
In this guide
- The GPUs are the expensive part, so feed them
- The six rules
Compare it with the next class up or down
The nearest classes we quote are the 8 GPU HGX B200 AI server and the 8 node HGX B300 AI pod, 64 GPUs. If neither is quite right, tell us the workload and we specify the system between them.
- (858) 613-1818
- sales@computermodules.com
- Monday to Friday, 8am to 5pm Pacific