NVIDIANew
GPU & AI Servers
NVIDIA DGX Rubin NVL8
Liquid-cooled DGX system for training, post-training and inference across every AI workload.
- GPUs: 8× NVIDIA Rubin
- Interconnect: Sixth-generation NVLink
- Cooling: Liquid-cooled
NVIDIA HGX Rubin NVL8
Eight Rubin GPUs on sixth-generation NVLink — the building block for partner servers from Supermicro, ASUS and others.
- GPUs8× NVIDIA Rubin
- InterconnectSixth-generation NVLink
- Host CPUNVIDIA Vera or x86 baseboards
- Form factorHGX baseboard for OEM systems
Overview
About the HGX Rubin NVL8
NVIDIA HGX Rubin NVL8 is an eight-GPU baseboard that links Rubin GPUs with sixth-generation NVLink. Server makers build it into their own systems, so you can choose the chassis, cooling and CPU platform that fits your data center.
NVIDIA rates HGX Rubin NVL8 at 10x the token throughput of HGX B200, and at matching training performance with 4x fewer GPUs. It is offered with x86 host CPUs (HGX Rubin NVL8) or with an NVIDIA Vera CPU (HGX Vera Rubin NVL8).
400 PFLOPS NVFP4
Extreme inference throughput for serving agentic and reasoning models.
2 TB HBM4
130 TB/s of GPU memory bandwidth, 2.1x more than the previous generation.
Sixth-generation NVLink
24 TB/s of NVLink Switch bandwidth across the eight GPUs.
MoE pre-training in 8 GPUs
4x more NVFP4 training FLOPS for mixture-of-experts models.
x86 or Vera host
Choose x86 baseboards from OEMs or the 88-core NVIDIA Vera CPU with up to 1.5 TB LPDDR5X.
1.8 TB/s networking
Bi-directional networking bandwidth for multi-node clusters.
Specifications
Technical Specifications
Platform
| Configuration | 8× NVIDIA Rubin SXM |
|---|---|
| Host CPU | x86 (OEM-defined) or single-socket NVIDIA Vera (88 Olympus cores, up to 1.5 TB LPDDR5X) |
| GPU memory | 2 TB HBM4 |
| GPU memory bandwidth | 130 TB/s |
| NVLink | Sixth generation, 24 TB/s NVLink Switch bandwidth |
| Networking bandwidth | 1.8 TB/s (bi-directional) |
| Inlet temperature | 45 °C |
Performance (8 GPUs)
| NVFP4 inference | 400 PFLOPS (sparse) |
|---|---|
| NVFP4 training | 270 PFLOPS (dense) |
| FP8/FP6 training | 130 PFLOPS (dense) |
| FP16/BF16 training | 32 PFLOPS (dense) |
| FP32 | 1,040 TFLOPS |
| FP64 | 265 TFLOPS |
Per Rubin GPU
| NVFP4 inference | 50 PFLOPS (sparse) |
|---|---|
| NVFP4 training | 35 PFLOPS (dense) |
| FP8/FP6 training | 17.5 PFLOPS (dense) |
| FP16/BF16 training | 4 PFLOPS (dense) |
| FP32 | FP64 | 130 TFLOPS | 33 TFLOPS |
| GPU memory | bandwidth | 288 GB HBM4 | 22 TB/s |
| NVLink bandwidth | 3.6 TB/s |
Source: official NVIDIA product information.
Specifications are set by the manufacturer and may change. Plan Man confirms the final configuration, availability and lead time in your quote.
Use cases
What it's built for
- Token-factory inference for agentic AI
- Mixture-of-experts pre-training
- LLM training and fine-tuning
- Data analytics
- HPC and scientific computing
Plan Man services
Supplied and supported in the UAE
- Requirement analysis, sizing and bill of materials
- Genuine supply with manufacturer warranty
- Rack integration, cabling and onsite deployment
- Power and liquid cooling planning for dense GPU racks
- After-sales support, RMA handling and upgrades from Dubai
NVIDIA
GPU & AI Servers
NVIDIA HGX B300
Blackwell Ultra 8-GPU platform for AI reasoning, available in air- and liquid-cooled partner servers.
- GPUs: 8× NVIDIA Blackwell Ultra
- GPU memory: 2.1 TB HBM3e total
- NVLink: 1.8 TB/s GPU-to-GPU
NVIDIA
GPU & AI Servers
NVIDIA HGX B200
The proven Blackwell 8-GPU platform for LLM training and high-volume inference.
- GPUs: 8× NVIDIA Blackwell
- GPU memory: 180 GB HBM3e per GPU
- NVLink: 1.8 TB/s GPU-to-GPU
Get a Quote
Tell us your workload and timeline for the NVIDIA HGX Rubin NVL8. We reply within one business day with a configuration, pricing and lead time.