(858) 613-1818Quote list 0

NVIDIA HGX Rubin NVL8

NVIDIA's next eight GPU baseboard with Rubin GPUs, 288 GB HBM4 each, sixth generation NVLink and 400 PFLOPS sparse NVFP4 per node. Register interest.

Family
Rubin
Edition
HGX 8 GPU baseboard, x86 or Vera CPU
Memory
288 GB HBM4 per GPU, 2 TB HBM4 per 8 GPU baseboard as published by NVIDIA
Memory bandwidth
22 TB/s per GPU, 130 TB/s per baseboard as published for the HGX Rubin NVL8
Interconnect
Sixth generation NVLink, 3.6 TB/s per GPU, 24 TB/s NVLink Switch bandwidth across 8 GPUs
Form factor
SXM module on the HGX Rubin NVL8 baseboard, paired with x86 CPUs or a single socket Vera CPU
TDP
Not yet published
Cooling
Not yet published
Best for
Eight GPU Rubin servers for standard racks, succeeding HGX B300
Availability
Coming soon

HGX Rubin NVL8 brings eight Rubin GPUs with sixth generation NVLink to the familiar HGX server form factor, with a choice of x86 processors or a single socket Vera CPU with 88 custom Olympus cores and up to 1.5 TB of LPDDR5X. NVIDIA publishes 288 GB of HBM4 per GPU for 2 TB per baseboard at 130 TB/s, 24 TB/s of NVLink Switch bandwidth, 1.8 TB/s of bi-directional networking, and 400 PFLOPS of sparse NVFP4 inference or 270 PFLOPS of dense NVFP4 training per node.

Against HGX B200, NVIDIA cites up to ten times the token throughput for reasoning inference and four times the NVFP4 training FLOPS. Chassis sizes, per GPU power, cooling options and channel availability are not yet published, so the HGX Rubin NVL8 server on this site is listed as coming soon. We are registering interest and will quote as soon as system partners release configurations; HGX B300 is the deliverable alternative today.

Systems that carry this GPU

Want this GPU in a different chassis?

We quote every class that carries it, from a single node to a pod. Tell us the workload and we propose the chassis.

Request a quotationView quote list