(858) 613-1818Quote list 0

1U dual socket AMD EPYC 9005 compute server

Dual AMD EPYC 9005 server in 1U with up to 384 cores, 24 DDR5 channels, 12 NVMe bays and 400 Gb/s networking, sized as the CPU tier of a GPU cluster.

Illustration of a 1U rackmount server
CPU
2 x AMD EPYC 9005 series, up to 192 cores per socket
Memory
24 x DDR5 RDIMM, 12 channels per socket, up to 3 TB
Storage
2 x M.2 boot, up to 12 x E3.S or U.2 NVMe
Networking
1 x 400 Gb/s ConnectX-7, 2 x 25 GbE onboard
Expansion
2 x PCIe Gen5 x16 slots, OCP 3.0 mezzanine
Power
Redundant PSUs, about 1 kW per node planning figure

Overview

This is the CPU node we place beside GPU servers to run data loading, tokenization, preprocessing, orchestration agents and the general purpose services a cluster needs. Two AMD EPYC 9005 processors give up to 384 cores and 24 DDR5 channels in one rack unit, with enough NVMe bays to stage datasets locally.

We size the CPU tier against the GPU count: about six threads and 24 GB of memory for every GPU in the cluster, so a 64 node HGX cluster with 512 GPUs needs roughly 3,072 threads and 12 TB of memory across its CPU tier. That works out to a handful of these nodes, which we add to cluster quotations by default.

Standard configuration is mirrored M.2 boot, up to twelve NVMe drives, a 400 Gb/s ConnectX-7 adapter for the fabric and dual 25 GbE for management and services. The node is also a strong general purpose server for databases, Kubernetes workers and virtualization at about 1 kW.

Use cases

  • Data preprocessing for training
  • Kubernetes and orchestration workers
  • Databases and vector stores
  • Virtualization hosts
  • General enterprise compute

Specifications

Compute
CPUs2 x AMD EPYC 9005 series, 16 to 192 cores per socket, up to 384 cores per node
System memory24 x DDR5 RDIMM, 12 channels per socket, 64 GB to 128 GB modules, up to 3 TB
Sizing rule6 CPU threads and 24 GB memory per cluster GPU served
Storage
Boot2 x M.2 NVMe, mirrored
DataUp to 12 x E3.S or U.2 NVMe, hot swap
RAIDSoftware RAID or optional hardware RAID
Networking
Fabric1 x 400 Gb/s ConnectX-7 or 2 x 200 Gb/s, PCIe Gen5
Frontend2 x 25 GbE onboard, OCP 3.0 slot for 100 GbE
ManagementDedicated 1 GbE BMC port, TPM 2.0
Physical
Form factor1U rackmount, air cooled
Power suppliesRedundant 1,200 W class PSUs, 100 to 240 V input
Planning powerAbout 1 kW per node

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

What's included

  • Rails, power cords and cable management arm
  • Burn in and acceptance test report
  • Operating system imaging and firmware baseline on request
  • 3 year warranty, options to 5 years with advance replacement

Compliance

This product is not a data center GPU system and does not require NVIDIA end user registration. Standard export documentation applies. Read about export compliance.

Compare it with the next class up or down

The nearest classes we quote are the 2U dual socket Xeon 6 server and the 1U EPYC 9005 control plane server. If neither is quite right, tell us the workload and we specify the system between them.

Talk to an engineerRequest a quotation