Skip to content
GIRIVON
Infrastructure+
Systems+
Components+
Computing Devices+
Solutions+
Compute Tools+
Knowledge+
NVIDIA DGX Station deskside AI supercomputer (official NVIDIA image)

NVIDIA · DGX Station — GB300 Grace Blackwell Ultra Desktop Superchip

NVIDIA DGX Station (GB300 Grace Blackwell Ultra)

Deskside AI supercomputer: Blackwell Ultra GPU with 252 GB HBM3e at 7.1 TB/s, 72-core Grace CPU with 496 GB LPDDR5X, 800 Gb/s ConnectX-8 networking.

₹1,67,47,713

Inclusive of 18% GST. Invoice details confirmed with your order.

Availability
Built to order
Delivery
Built to order
SKU
NV-DGX-STATION-GB300
Warranty
Per manufacturer
Request quote

Key specifications

GPU count
1
Memory per GPU
252 GB
Memory
496 GB

NVIDIA DGX Station is a deskside AI supercomputer built on the GB300 Grace Blackwell Ultra Desktop Superchip: one Blackwell Ultra GPU with 252 GB of HBM3e ECC memory at 7.1 TB/s, connected over 900 GB/s NVLink-C2C to a 72-core Grace CPU with 496 GB of LPDDR5X ECC memory. A ConnectX-8 SuperNIC provides up to 800 Gb/s networking, and one additional NVIDIA RTX PRO Blackwell GPU can be fitted for visualisation and simulation. Total system power is 1,600 W on a 20 A circuit.

DGX Station is built by NVIDIA's system partners. Price is based on the starting price an NVIDIA partner (Exxact, Valence VWS-158270643) publishes for its DGX Station build (USD); options are quoted on request. Specifications are NVIDIA's published figures (Tensor Core figures with sparsity unless noted).

Full specifications

Accelerator

GPU1x NVIDIA Blackwell UltraVendor-reported
GPUs1Vendor-reported
GPU memory252 GBVendor-reported
GPU memory typeHBM3e ECCVendor-reported
GPU memory bandwidth7100 GB/sVendor-reported
Multi-Instance GPUUp to 7 instancesVendor-reported

Compute

CPU1x NVIDIA Grace 72-core Neoverse V2Vendor-reported
CPU cores72Vendor-reported

Memory

CPU memory496 GBVendor-reported
CPU memory typeLPDDR5X ECC, 396 GB/sVendor-reported
ECC memoryYesVendor-reported

Networking

CPU–GPU interconnectNVLink-C2C, 900 GB/sVendor-reported
Networking (peak)800 Gb/sVendor-reported
Network adapterNVIDIA ConnectX-8 SuperNIC (Ethernet)Vendor-reported
Ethernet ports2Vendor-reported
Port type2x QSFP112 (400 Gb/s per port)Vendor-reported

Performance

FP4 Tensor Core (sparse | dense)20 | 15 PFLOPSVendor-reported
FP8 / FP6 Tensor Core (sparse)10 PFLOPSVendor-reported
FP16 / BF16 Tensor Core (sparse)5 PFLOPSVendor-reported
FP3280 TFLOPSVendor-reported
FP64 / FP64 Tensor Core1.3 TFLOPSVendor-reported

Storage

Storage slots (M.2 Gen 5)4Vendor-reported

Power

Total system power1600 WVendor-reported
Electrical circuit20 A circuit requiredVendor-reported

Software

Operating systemUbuntu with NVIDIA AI Developer ToolsVendor-reported

Workload suitability

  • calculated LLM inference: about 113B parameters per GPU in BF16, or 403B in 4-bit (weights only, 252 GB per GPU).
  • calculated Fine-tuning: QLoRA of models up to about 294B parameters per GPU; full fine-tuning (≈16 bytes/parameter) of about 12.6B parameters across all GPUs.
  • calculated Rendering: scenes up to roughly 252 GB must fit in each GPU's memory for GPU renderers.
  • calculated Parallel CPU work: 72 cores, 7 GB RAM per core.
  • calculated Virtualisation: roughly 31 VMs of 4 vCPU / 16 GB at a 3:1 vCPU ratio.

Performance & benchmarks

No measured benchmarks are published for this product yet. GIRIVON does not publish estimated figures as benchmarks; use the workload performance estimator with your own measurements.

Power & thermal

Rated power
1,600 W
Vendor-reported
Heat at rated power
5,459 BTU/h
calculated

Compatibility

The compatibility engine validates this product against other components using its structured attributes (17 recorded).

Warranty & support

Manufacturer warranty applies. For installation, RMA and technical help, contact GIRIVON support.

Documentation

Reviews

No reviews yet. Reviews are moderated before publication.

    Write a review
    Rating

    Related products