(858) 613-1818Quote list 0
Coming soon

8 GPU HGX Rubin NVL8 AI server

Next generation eight GPU server on the NVIDIA HGX Rubin NVL8 baseboard with 288 GB HBM4 per GPU and sixth generation NVLink. Register interest now.

GPU: NVIDIA HGX Rubin NVL8

Illustration of a 10U HGX GPU server with the lid open
GPUs
8 x NVIDIA Rubin, 288 GB HBM4 each, sixth generation NVLink; 2 TB per node as published for HGX Vera Rubin NVL8
CPU
x86 (AMD EPYC 9005 or Intel Xeon 6) or single socket NVIDIA Vera, per NVIDIA's published options
Memory
DDR5 with x86 hosts, up to 1.5 TB LPDDR5X with Vera
Storage
Not yet published; expect mirrored boot and NVMe data bays as in HGX B300
Networking
8 x 800 Gb/s east west ports, ConnectX-8 class; 1.8 TB/s bi-directional per NVIDIA
Power
Not yet published

Overview

HGX Rubin NVL8 is NVIDIA's next eight GPU baseboard and the successor to HGX B300 in standard racks, with a choice of x86 processors or a single socket Vera CPU. The node level figures NVIDIA has published so far, 2 TB of HBM4 at 130 TB/s, sixth generation NVLink at 3.6 TB/s per GPU and 400 PFLOPS of sparse NVFP4 inference, are as published for HGX Vera Rubin NVL8; the column for the x86 HGX Rubin NVL8 is not yet published. Each Rubin GPU carries 288 GB of HBM4 at 22 TB/s.

Chassis sizes, per node power, cooling options and reseller delivery dates have not yet been published, so this server is listed as coming soon and we are not taking firm orders. What we can do now is register your interest, reserve a position in our allocation queue, and review your data center against the HGX B300 requirements, since Rubin systems will need at least that much power and cooling.

If your project cannot wait, the air cooled and liquid cooled HGX B300 servers are deliverable today and run the same CUDA software, so a cluster can start on Blackwell Ultra and add Rubin nodes as they ship. Ask us for a roadmap conversation and we will keep you updated as NVIDIA releases specifications.

Use cases

  • Planning 2026 and 2027 training clusters
  • Roadmap alignment for AI factories
  • Reasoning model inference at scale

Specifications

Compute
GPUs8 x NVIDIA Rubin SXM on the HGX Rubin NVL8 baseboard
GPU memory288 GB HBM4 per GPU; 2 TB and 130 TB/s per node as published for HGX Vera Rubin NVL8, x86 HGX Rubin NVL8 figures not yet published
GPU interconnectSixth generation NVLink, 3.6 TB/s per GPU, 24 TB/s NVLink Switch bandwidth
NVFP4 performance400 PFLOPS sparse inference, 270 PFLOPS dense training per node as published for HGX Vera Rubin NVL8; x86 variant not yet published
CPUsx86 dual socket or 1 x NVIDIA Vera with 88 Olympus cores
Networking
GPU fabric8 x 800 Gb/s ports, ConnectX-8 class, InfiniBand or Ethernet
FrontendNot yet published; BlueField-3 or BlueField-4 class expected
ManagementDedicated BMC port
Physical
Form factorNot yet published; 10U air cooled and liquid cooled chassis expected from system partners
Planning powerNot yet published
CoolingNot yet published

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

What's included

  • Expression of interest and queue position
  • Site readiness review against HGX B300 requirements
  • Specification updates as NVIDIA publishes them
  • Firm quotation when configurations are released

Compliance

Compare it with the next class up or down

The nearest classes we quote are the 8 GPU HGX B300 AI server, air cooled and the NVIDIA Vera Rubin NVL72 rack. If neither is quite right, tell us the workload and we specify the system between them.

Talk to an engineerRequest a quotation