NVIDIA DGX Station™ —— The Ultimate Deskside AI Supercomputer
Category | Specification |
|---|---|
Superchip | NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip™ |
GPU | 1× NVIDIA Blackwell Ultra GPU · latest-generation Tensor Cores · NVFP4/FP4 support |
CPU | 1× NVIDIA Grace 72-Core (Arm Neoverse V2) |
AI Compute | FP4 Tensor Core: up to 153 PFLOPS (sparse peak, official spec) |
FP8 / FP6 Tensor Core: 10 PFLOPS · FP16 / BF16: 5 PFLOPS · TF32: 2.5 PFLOPS · FP32: 80 TFLOPS · FP64: 1.3 TFLOPS · INT8: 330 TOPS | |
Marketing headline "up to 20 petaFLOPS FP4" = NVFP4 aggregate system figure; spec-table peak is 153 PFLOPS — state the two separately. † | |
GPU Memory | 252 GB HBM3e ECC · 7.1 TB/s bandwidth |
CPU Memory | 496 GB LPDDR5X ECC · 396 GB/s bandwidth |
Coherent Memory | 748 GB CPU+GPU coherent memory pool (NVLink-C2C) |
Interconnect | NVIDIA NVLink®-C2C · 900 GB/s chip-to-chip bandwidth |
Multi-Instance GPU | Up to 7 MIG instances, each with independent HBM, cache & compute cores |
Storage | 4× M.2 Gen 5 slots |
Networking | NVIDIA ConnectX®-8 SuperNIC · up to 800 Gb/s Ethernet · 2× QSFP112 (400 Gb/s per port) · 1× RJ-45 10 GbE · 1× RJ-45 1 GbE (BMC) |
Link up to 2 DGX Stations for larger models / network-accelerated transfers | |
Front I/O | 2× USB Type-C (USB 3.1) · 2× USB Type-A (USB 3.1) · Audio |
Rear I/O | 4× USB Type-A · USB Micro-B (BMC) · Mini DisplayPort 5 (BMC only, for system management) · Audio |
Optional GPU | Up to 1 additional NVIDIA RTX PRO™ Blackwell GPU: RTX PRO 6000 Workstation Edition · RTX PRO 6000 Blackwell Max-Q Workstation Edition · RTX PRO 4000 Blackwell SFF Edition · RTX PRO 2000 Blackwell |
OS | Ubuntu with NVIDIA AI Developer Tools (Linux) · DGX Station for Windows (Windows, Q4; WSL-supported Linux workloads) |
Software | Preconfigured DGX OS + CUDA-X, NIM microservices (NIM support planned June, production-branch support planned October, per official footnote), NVIDIA AI Enterprise license available |
Manageability | BMC out-of-band telemetry · Redfish® API · NVIDIA Data Center GPU Manager (DCGM) · Hardware root of trust · Enterprise secure boot |
System Power | 1,600 W · 20 A circuit required |
Dimensions / Weight / Acoustics | Not published on the current spec page † |
One memory pool, no PCIe bottleneck. The Blackwell Ultra GPU and 72-core Grace CPU talk over NVLink-C2C at 900 GB/s, so massive models and long-context agents live in a single coherent 748 GB pool instead of bouncing across a PCIe bus.
A personal AI cloud. Slice the GPU into up to 7 MIG instances — one desk, a whole team, simultaneous training and inference jobs, each isolated with its own HBM and compute. Or run hundreds of agents at once.
Train locally, scale globally. Identical software stack to NVIDIA AI factories: PyTorch, Jupyter, Ollama, NIM, Blueprints. Prototype here, ship to DGX Cloud or any accelerated data center with no re-platforming.
Built for IT. BMC + Redfish + DCGM telemetry, hardware root of trust and secure boot put GB300 in the same fleet-management world as your servers — not a boutique workstation.
Windows, too. DGX Station for Windows brings the same GB300 superchip into the enterprise Windows ecosystem, with OpenShell agent sandboxes on Windows containment primitives and Linux workloads via WSL.