DGX Station
The NVIDIA DGX Station is a deskside AI supercomputer designed for developers, researchers, and data scientists who need to build, fine-tune, and run large AI models locally.
Publisher review
The NVIDIA DGX Station is a deskside AI supercomputer designed for developers, researchers, and data scientists who need to build, fine-tune, and run large AI models locally. Powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, it delivers up to 20 petaFLOPS of AI compute performance and 748 GB of coherent memory, enabling work with models up to 1 trillion parameters. The system is preconfigured with Ubuntu and NVIDIA AI Developer Tools, including CUDA-X libraries, for a full-stack AI development environment. A separate Windows edition targets enterprise users who want to run always-on AI agents integrated with existing workflows.
The DGX Station's architecture pairs a Blackwell Ultra GPU with a high-performance Grace CPU via the NVLink-C2C interconnect, providing 288 GB of HBM3e GPU memory with 8 TB/s bandwidth and 496 GB of CPU LPDDR5X memory at 396 GB/s. It includes NVIDIA Blackwell Generation Tensor Cores with 4-bit floating point (NVFP4) AI, which increases the performance and size of models that memory can support while maintaining accuracy. The NVIDIA ConnectX-8 SuperNIC provides up to 800 Gb/s network connectivity and supports linking two DGX Stations to scale model capacity. This configuration allows rapid development, fine-tuning, and inference of large models, with seamless deployment to data center or cloud.
Compared to the smaller DGX Spark (priced around $4,300, with 128 GB unified memory and 273 GB/s bandwidth), the DGX Station offers roughly 20x the AI compute and 2.25x the GPU memory, but at a significantly higher cost. The DGX Spark's 273 GB/s memory bandwidth has been criticized as lower than Apple's M4 Max (~400 GB/s) and NVIDIA's own RTX 5090 (1,700 GB/s), though the DGX Station's 8 TB/s GPU memory bandwidth eliminates that bottleneck. The DGX Station targets workloads that the Spark cannot handle, such as trillion-parameter model training and multi-agent systems, while the Spark is suited for prototyping with models up to 200 billion parameters.
The primary trade-offs are cost and specialization. The DGX Station is estimated to cost between $40,000 and $60,000, making it a major capital investment for teams that need local, private AI compute. Its specialized DGX OS and software stack may limit versatility compared to general-purpose workstations. The 273 GB/s memory bandwidth of the smaller DGX Spark is a real limitation for token generation speed in LLMs, though the DGX Station's HBM3e memory avoids this issue. Real-world deployment reports for both systems remain limited as of mid-2025, and the high price of the Station may be hard to justify for teams that could use cloud instances instead.
How it works
-
GB300 Grace Blackwell Ultra Superchip
Combines a Blackwell Ultra GPU with a Grace CPU via NVLink-C2C for high-bandwidth, low-latency communication between processor and memory.
-
20 petaFLOPS AI compute
Delivers up to 20 petaFLOPS of AI performance, enabling training and inference of models with up to 1 trillion parameters.
-
748 GB coherent memory
Provides 748 GB of unified memory, including 288 GB HBM3e GPU memory at 8 TB/s bandwidth and 496 GB CPU LPDDR5X at 396 GB/s.
-
NVFP4 Tensor Cores
Blackwell Generation Tensor Cores support 4-bit floating point (NVFP4) AI, increasing model size and performance while maintaining accuracy.
-
ConnectX-8 SuperNIC networking
Offers up to 800 Gb/s network connectivity and supports linking two DGX Stations to scale model capacity and performance.
-
Preconfigured AI software stack
Comes with Ubuntu and NVIDIA AI Developer Tools, including CUDA-X libraries, for a ready-to-use AI development environment.
-
Deskside form factor
Compact design fits on a desk, allowing local development and inference without requiring data center space or cloud resources.
Strengths and trade-offs
Strengths
- Delivers up to 20 petaFLOPS of AI compute, enabling training and inference of models with up to 1 trillion parameters locally.
- Provides 288 GB of HBM3e GPU memory with 8 TB/s bandwidth, eliminating memory bandwidth bottlenecks common in smaller systems.
- Supports linking two DGX Stations via ConnectX-8 SuperNIC at 800 Gb/s to scale model capacity and performance further.
- Includes NVIDIA Blackwell Generation Tensor Cores with NVFP4 precision, increasing model size and performance while maintaining high accuracy.
Trade-offs
- Estimated price of $40,000 to $60,000 makes it a major capital investment, potentially hard to justify versus cloud instances.
- Specialized DGX OS may restrict versatility compared to general-purpose workstations for non-AI workloads.
- Real-world deployment reports are limited as of mid-2025, making it difficult to assess performance in production environments.
- The smaller DGX Spark's 273 GB/s memory bandwidth is lower than Apple's M4 Max (~400 GB/s) and NVIDIA's RTX 5090 (1,700 GB/s), though the Station's HBM3e avoids this issue.
Pricing context
DGX Station estimated at $40,000–$60,000; DGX Spark starts at $2,999 (ASUS Ascent GX10) or $4,300 (NVIDIA direct).
Getting started with DGX Station
-
Unbox and power on
Unpack the DGX Station and place it on a stable desk near a power outlet. Connect the power cable and press the power button. Wait for the system to boot into the preinstalled Ubuntu environment.
-
Connect to network
Attach an Ethernet cable to the ConnectX-8 SuperNIC port for high-speed networking. Alternatively, use the built-in Wi-Fi for initial setup. Configure network settings in Ubuntu to ensure internet access.
-
Install NVIDIA AI tools
Open a terminal and update the package list with 'sudo apt update'. Install the NVIDIA AI Developer Tools by running 'sudo apt install nvidia-ai-workbench'. This includes CUDA-X libraries and other dependencies.
-
Load a large model
Download a pre-trained model up to 1 trillion parameters from NVIDIA NGC or Hugging Face using the command line. For example, run 'ngc registry model download nvidia/llama-3.1-405b' to pull the model into the local workspace.
-
Run inference locally
Use the NVIDIA AI Workbench to launch an inference script. Execute 'python run_inference.py --model_path ./llama-3.1-405b --prompt "Explain AI"' to test the model. Monitor GPU utilization with 'nvidia-smi' to verify performance.
Frequently Asked Questions
What is the NVIDIA DGX Station?
The NVIDIA DGX Station is a deskside AI supercomputer for developers and researchers. It uses the GB300 Grace Blackwell Ultra Superchip to deliver up to 20 petaFLOPS of AI compute and 748 GB of coherent memory, enabling local work with models up to 1 trillion parameters.
How much does the DGX Station cost?
The DGX Station is estimated to cost between $40,000 and $60,000. This makes it a major capital investment for teams needing local, private AI compute. In comparison, the smaller DGX Spark starts at around $2,999 from ASUS or $4,300 directly from NVIDIA.
What are the key specs of the DGX Station?
The DGX Station features 20 petaFLOPS AI compute, 748 GB coherent memory (288 GB HBM3e GPU memory with 8 TB/s bandwidth and 496 GB CPU LPDDR5X), NVFP4 Tensor Cores, and up to 800 Gb/s networking via ConnectX-8 SuperNIC. It runs Ubuntu with NVIDIA AI tools.
How does the DGX Station compare to the DGX Spark?
The DGX Station offers roughly 20 times the AI compute and 2.25 times the GPU memory of the DGX Spark. While the Spark handles models up to 200 billion parameters, the Station supports trillion-parameter models. However, the Station costs significantly more, estimated at $40,000–$60,000.
Who should buy the DGX Station?
The DGX Station targets developers, researchers, and data scientists who need to build, fine-tune, and run large AI models locally. It suits teams requiring private compute for trillion-parameter models or multi-agent systems, where cloud instances may not be ideal due to cost or data sensitivity.
What are the main trade-offs of the DGX Station?
The main trade-offs are its high cost, estimated at $40,000–$60,000, and specialized DGX OS that may limit versatility for non-AI tasks. Real-world deployment reports remain limited as of mid-2025, and cloud instances might be more cost-effective for some teams.
Alternatives
How DGX Station compares
Direct head-to-head against 3 competitors. Picked by 7wData.
DGX Station
- Pricing
- DGX Station estimated at $40,000–$60,000; DGX Spark starts at $2,999 (ASUS Ascent GX10) or $4,300 (NVIDIA direct).
- Target
- The NVIDIA DGX Station is a deskside AI supercomputer designed for developers, researchers, and data scientists who need to build, fine-tune, and run large AI
- Strength
- Delivers up to 20 petaFLOPS of AI compute, enabling training and inference of models with up to 1 trillion parameters locally.
- Watch for
- Estimated price of $40,000 to $60,000 makes it a major capital investment, potentially hard to justify versus cloud instances.
AMD Strix Halo
- Pricing
- $2,199-$2,950
- Target
- Developers needing unified memory
- Deployment
- Linux/Windows
- Strength
- Half the price of DGX Spark
- Watch for
- ROCm ecosystem lacks CUDA depth
Apple Mac Studio M4 Ultra
- Pricing
- $4,400+
- Target
- Bandwidth-intensive workloads
- Deployment
- macOS only
- Strength
- 819 GB/s memory bandwidth
- Watch for
- Limited to macOS ecosystem
NVIDIA RTX 5090 Custom Build
- Pricing
- $2,800-$3,200
- Target
- High-performance AI workloads
- Deployment
- Linux/Windows
- Strength
- 1,792 GB/s memory bandwidth
- Watch for
- High power consumption (725W+)
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.