NVIDIA DGX Spark and DGX Station Personal AI Computers

NVIDIA Releases the Personal AI Computers DGX Spark and DGX Station Powered by NVIDIA.

According to Grace Blackwell, desktop supercomputers from prominent computer manufacturers like ASUS, Dell Technologies, HP, and Lenovo give developers, researchers, and data scientists access to faster AI. The NVIDIA Grace Blackwell platform-powered NVIDIA DGX personal AI supercomputers were announced today by NVIDIA.

AI developers, researchers, data scientists, and students may prototype, refine, and infer huge models on desktops with the help of DGX Spark, formerly Project DIGITS, and DGX Station, a new high-performance NVIDIA Grace Blackwell desktop supercomputer driven by the NVIDIA Blackwell Ultra platform. These models can be deployed on NVIDIA DGX Cloud or any other accelerated cloud or data centre infrastructure, or users can run them locally.

The Grace Blackwell architecture’s power, which was previously limited to the data centre, is now accessible on desktops to DGX Spark and DGX Station. ASUS, Dell, HP Inc., and Lenovo are among the international system manufacturers working on DGX Spark and DGX Station.

AI has revolutionised the entire computing stack. It makes sense that a new class of machines would appear that are intended to run AI-native apps and be used by AI-native developers. “Cloud services, desktop, and edge applications can all benefit from AI with these new DGX personal AI computers.”

Igniting Innovation With DGX Spark

With its enormous performance and capabilities, DGX Spark is the smallest AI supercomputer in the world, enabling millions of academics, data scientists, robotics developers, and students to push the limits of generative and physical AI.

The NVIDIA GB10 Grace Blackwell Superchip, which is tailored for a desktop form factor, is the central component of DGX Spark. For fine-tuning and inference with the newest AI reasoning models, such as the NVIDIA Cosmos Reason world foundation model and the NVIDIA GR00T N1 robot foundation model, the GB10 boasts a potent NVIDIA Blackwell GPU with fifth-generation Tensor Cores and FP4 support, capable of delivering up to 1,000 trillion operations per second of AI compute.

The GB10 Superchip offers a CPU+GPU-coherent memory model with five times the bandwidth of fifth-generation PCIe to NVIDIA NVLink-C2C connection technology. In order to maximize performance for memory-intensive AI development workloads, this enables the superchip to access data between a GPU and CPU.

Prototyping, optimizing, and iterating workflows is now easier than ever to NVIDIA’s full-stack AI platform, which allows DGX Spark users to transfer their models from their desktops to DGX Cloud or any accelerated cloud or data centre infrastructure with almost no code changes.

Read more on NVIDIA Blackwell Ultra DGX SuperPOD: Agentic AI at Scale

Full Speed Ahead With DGX Station

For AI development, NVIDIA DGX Station gives desktops data center-level performance. With an enormous 784GB of coherent memory space, the DGX Station the first desktop machine constructed with the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip is able to speed up demanding training and inferencing tasks. Best-in-class system connectivity and performance are provided by the GB300 Desktop Superchip’s NVIDIA Blackwell Ultra GPU, which has the newest Tensor Cores and FP4 accuracy, and is coupled to a powerful NVIDIA Grace CPU via NVLink-C2C.

The NVIDIA ConnectX-8 SuperNIC, designed to boost hyperscale AI computing workloads, is another aspect of DGX Station. The ConnectX-8 SuperNIC provides incredibly fast and efficient network access with capability for networking at up to 800Gb/s. This allows for network-accelerated data transfers for AI workloads and high-speed connectivity of numerous DGX Stations for even bigger workloads.

Teams can obtain remarkable desktop AI development performance by combining the NVIDIA CUDA-X AI platform with these cutting-edge DGX Station capabilities.

Additionally, the NVIDIA AI Enterprise software platform provides highly optimized, enterprise-supported inference microservices that are simple to deploy, giving users access to NVIDIA NIM microservices.

Accessibility

DGX Spark system reservations are now being accepted. Later this year, manufacturing partners ASUS, BOXX, Dell, HP, Lambda, and Supermicro are anticipated to make DGX Station available.

DGX Station Price

To help lower the expenses of installing these systems on a small scale, Nvidia also provides Station rental at about US$9000 per month through partners in the US (rentacomputer.com) and Europe (iRent IT Systems). The DGX Station A100 320G costs $149,000, while the 160G variant costs $99,000.

Overview

NVIDIA DGX Spark

The NVIDIA DGX Spark, which is powered by the NVIDIA GB10 Grace Blackwell Superchip, offers 1000 AI TOPS of AI performance in a small, power-efficient package. The most recent generation of reasoning AI models from DeepSeek, Meta, Google, and others may be prototyped, refined, and inferred locally with up to 200 billion parameters using the NVIDIA AI software stack that comes preloaded and 128GB of memory. These models can then be smoothly deployed to the data centre or cloud.

Features

• NVIDIA Software Technologies for GPU, CPU, Networking, and AI: NVIDIA GB10 Superchip: With the NVIDIA Grace Blackwell architecture, achieve up to 1000 AI TOPS of AI performance at FP4 accuracy.

• 128 GB of Coherent Unified System Memory: Use a significant amount of unified system memory on your desktop to run AI development and testing workloads with AI models up to 200 billion parameters.

• High-performance NVIDIA ConnectX networking: Two NVIDIA DGX Spark systems can be connected using NVIDIA Connect-X networking to work with AI models that have up to 405 billion parameters.

• NVIDIA AI Software Stack: For generative AI tasks, use a full-stack solution that includes pretrained models, frameworks, tools, and libraries.

Accelerate All AI Workloads

NVIDIA DGX Spark is perfect for workloads including data scientists, researchers, and AI developers because it provides the power of an AI supercomputer in a desktop-friendly size.

Making prototypes: Create, evaluate, and verify AI applications and models. NVIDIA DGX Spark offers developers the platform to design, test, and validate AI models as well as AI-enhanced apps and solutions through the NVIDIA AI software stack. Work can be effortlessly moved to NVIDIA DGX cloud or other NVIDIA-accelerated data centres or cloud infrastructures for final tuning or deployment.

Adjusting: Adjust up to 70 billion parameters in AI models. By fine-tuning on NVIDIA DGX Spark, pre-trained models can perform better. Customize AI models and solutions for particular requirements and use cases by fine-tuning models up to 70 billion parameters using 128GB of unified system memory.

Inference: Use AI models with up to 200 billion parameters for testing, validation, and inference. Together with 128GB of system memory, fifth-generation Tensor Cores with FP4 support provide up to 1,000 TOPS of AI computation performance and speed up the inference of cutting-edge AI models for testing, validating, and deploying from your NVIDIA DGX Spark.

Information Science: Use NVIDIA RAPIDS to speed up data science workflows from start to finish. With zero-code-change accelerators and well-known APIs, NVIDIA RAPIDS on NVIDIA DGX Spark streamlines the development process from data preparation to model training, inference, and deployment. This allows for the quick acceleration of current workloads or the rapid development of use-case solutions, as well as the easy scaling out to data centres or the cloud.

Applications at the Edge: Use NVIDIA AI frameworks, such as Metropolis, Isaac, and many more, to create edge applications. A great platform for creating computer vision, smart city, and robotics solutions is offered by DGX Spark. frameworks from NVIDIA, including as Holoscan, Metropolis, and Isaac. Developers may quickly create edge applications by utilising the capabilities of NVIDIA DGX Spark to these frameworks and tools.

NVIDIA DGX Spark Specifications