(858) 613-1818Quote list 0

4 GPU RTX 5090 AI server, 4U

Four GeForce RTX 5090 GPUs with 128 GB of GDDR7 in a 4U air cooled server on EPYC 9005 or Xeon 6, the lowest cost per GPU for inference and fine tuning.

GPU: NVIDIA GeForce RTX 5090

Illustration of a 4U GPU server
GPUs
4 x NVIDIA GeForce RTX 5090, 32 GB GDDR7 each, 128 GB total, 21,760 CUDA cores each
CPU
1 or 2 x AMD EPYC 9005 or Intel Xeon 6
Memory
Up to 1.5 TB DDR5 RDIMM
Storage
2 x M.2 boot, up to 8 x U.2 NVMe
Networking
2 x 25 GbE onboard, optional 100 or 200 Gb/s adapter
Power
Redundant PSUs, about 3.5 kW per node planning figure

Overview

The 4 GPU RTX 5090 server puts four GeForce RTX 5090 cards, 128 GB of GDDR7 in total, into a 4U air cooled chassis on a single or dual socket AMD EPYC 9005 or Intel Xeon 6 host. It is the lowest cost per GPU in our catalog for inference, fine tuning of small and mid size models, image and video generation, rendering farms and development clusters.

Each card has 21,760 CUDA cores, fifth generation Tensor cores with FP4 support and 1.8 TB/s of memory bandwidth, so one node serves four independent replicas of a model up to the 30B class at FP4, or one larger model split across the cards over PCIe. Hardware AV1 encode and decode on every GPU suits media generation pipelines. There is no NVLink and no ECC on the GPU memory.

GeForce cards are consumer products: NVIDIA does not license the GeForce driver for data center deployment, and the GPU memory is not ECC protected. The server suits buyers who accept those terms for a lab, an office server room or a development cluster, and we ask you to confirm the intended use on the quotation. For ECC, MIG and professional drivers, the 4 GPU 4U PCIe server with RTX PRO 6000 Blackwell Server Edition is the step up.

Use cases

  • Low cost inference
  • Fine tuning small and mid size models
  • Image and video generation
  • Render farms
  • Development clusters

Specifications

Compute
GPUs4 x NVIDIA GeForce RTX 5090, 575 W each, spaced for airflow
GPU memory32 GB GDDR7 per GPU, 512 bit, 1.8 TB/s, no ECC
GPU interconnectPCIe Gen5 x16 per GPU, no NVLink
CPUs1 x or 2 x AMD EPYC 9005 series or Intel Xeon 6
System memory12 or 24 x DDR5 RDIMM, up to 1.5 TB
Storage
Boot2 x M.2 NVMe, mirrored
DataUp to 8 x U.2 NVMe, hot swap
RAIDSoftware RAID, optional hardware RAID or HBA
Networking
Onboard2 x 25 GbE
Optional1 x 100 or 200 Gb/s ConnectX-7 adapter
ManagementDedicated 1 GbE BMC port, TPM 2.0
Physical
Form factor4U rackmount, air cooled
Power suppliesRedundant 2,000 W class PSUs, 200 to 240 V input
Planning powerAbout 3.5 kW per node
Driver and licensingGeForce driver; the buyer confirms the intended use on the quotation

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

What's included

  • Rails and power cords
  • Burn in with all GPUs under sustained load and acceptance report
  • Operating system imaged with NVIDIA drivers and CUDA toolkit 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 8 GPU RTX 5090 AI server, 4U and the 4 GPU PCIe AI server, 4U. If neither is quite right, tell us the workload and we specify the system between them.

Talk to an engineerRequest a quotation