NVIDIA DGX Station — GB300 Grace Blackwell Ultra Desktop AI Supercomputer

Specifications

Overview

NVIDIA DGX Station is a deskside personal AI supercomputer built on the GB300 Grace Blackwell Ultra Desktop Superchip. It puts 252 GB of HBM3e GPU memory running at 7.1 TB/s alongside 496 GB of LPDDR5X CPU memory at 396 GB/s, joined by a 900 GB/s NVLink-C2C link — a coherent memory pool sized for frontier models that would otherwise require a datacenter allocation.

Compute is rated up to 20 PFLOPS at FP4 with sparsity, 10 PFLOPS at FP8/FP6, 5 PFLOPS at FP16/BF16, 2.5 PFLOPS TF32, 330 TOPS INT8, 80 TFLOPS FP32 and 1.3 TFLOPS FP64, with seven MIG instances for workload isolation and seven NVDEC / seven nvJPEG decoder blocks for media pipelines.

Connectivity comes from an NVIDIA ConnectX-8 SuperNIC at up to 800 Gb/s Ethernet, two QSFP112 ports at 400 Gb/s each, a 10 GbE RJ45 and a dedicated 1 GbE BMC port. Storage is served by four M.2 Gen 5 slots, with three PCIe Gen5 expansion slots and support for RTX PRO 6000 Blackwell, Max-Q, 4000 SFF and 2000 add-in cards. Total system power is 1,600 W on Ubuntu with NVIDIA AI Developer Tools.

Key Benefits

252 GB HBM3e at 7.1 TB/s — datacenter-class GPU memory on a desk. 496 GB LPDDR5X plus 900 GB/s NVLink-C2C gives a coherent pool for frontier-scale models. 20 PFLOPS FP4 with a dedicated transformer engine. 7 MIG instances for safe multi-workload sharing on one system. 800 Gb/s ConnectX-8 networking lets several stations behave like a small cluster. Available from seven OEMs — ASUS, Boxx, Dell, GIGABYTE, HP, MSI and Supermicro.

Applications

Local training and fine-tuning of frontier open-source models, inference at the desk for large-context agents, model development and evaluation without datacenter queueing, multi-station clustered research environments, and regulated or air-gapped AI development where data cannot leave the room.

Related Products