Liquid-cooled DGX system for training, post-training and inference across every AI workload.
NVIDIA DGX Rubin NVL8 is a turnkey eight-GPU DGX system for agentic AI and reasoning, covering training, post-training and inference on one platform. It delivers 400 petaFLOPS of NVFP4 performance with 2.3 TB of GPU memory.
The system is managed with NVIDIA Mission Control and fits into DGX BasePOD and DGX SuperPOD reference architectures, so it can start as a single node and grow into a cluster.
Exceptional inference performance with ultra-high memory bandwidth for models and agents.
The Rubin architecture includes engines that accelerate multi-agent reasoning workflows.
28.8 TB/s of NVLink Switch bandwidth for peer-to-peer GPU communication.
Full-lifecycle management from configuration to cluster operations and workload orchestration.
Eight ConnectX-9 VPI adapters plus BlueField-4 DPUs for storage and management traffic.
Integrates with DGX BasePOD and DGX SuperPOD reference designs.
| GPUs | 8× NVIDIA Rubin |
|---|---|
| Total GPU memory | 2.3 TB |
| GPU memory bandwidth | 176 TB/s |
| NVLink Switch bandwidth | 28.8 TB/s total |
| NVFP4 inference | 400 PFLOPS |
|---|---|
| NVFP4 training | 280 PFLOPS (dense) |
| FP8/FP6 training | 140 PFLOPS (dense) |
| CPUs | 2× Intel Xeon 6776P |
|---|---|
| Networking | 8× OSFP ports serving 8× single-port NVIDIA ConnectX-9 VPI, up to 800 Gb/s InfiniBand and Ethernet |
| DPU | 2× 400G QSFP112 ports on NVIDIA BlueField-4 |
| System power | ~24 kW |
| Software | NVIDIA DGX OS, Ubuntu, Red Hat Enterprise Linux, Rocky Linux; NVIDIA Mission Control |
Eight Rubin GPUs on sixth-generation NVLink — the building block for partner servers from Supermicro, ASUS and others.
Blackwell Ultra 8-GPU platform for AI reasoning, available in air- and liquid-cooled partner servers.
The proven Blackwell 8-GPU platform for LLM training and high-volume inference.
Tell us your workload and timeline for the NVIDIA DGX Rubin NVL8. We reply within one business day with a configuration, pricing and lead time.