Nvidia Starts Shipping ‘World’s Smallest AI Supercomputer’

Nvidia recently began shipping the DGX Spark, a compact desktop system it calls the "world's smallest AI supercomputer."

A DGX Spark workstation setup. 

Designed to bring high-end AI development out of data centers and into labs and offices, the DGX Spark integrates Nvidia's full AI software stack with the latest Grace Blackwell architecture to deliver up to one petaFLOP of AI performance in a 1.2-kg form factor.

A Closer Look at DGX Spark 

Nvidia built DGX Spark (datasheet linked) to overcome some of the shortcomings of conventional workstations in memory bandwidth and generative AI compute. The system enables developers to fine tune models with up to 70 billion parameters and run inference on models as large as 200 billion parameters locally. When paired via its built-in ConnectX-7 200 Gb/s networking interface, two units can support inference for models with up to 405 billion parameters.

The core of DGX Spark is Nvidia's GB10 Grace Blackwell Superchip. The processor tightly integrates a 20-core Grace CPU and a Blackwell GPU using NVLink-C2C, which provides 5x the bandwidth of PCIe Gen 5. The system’s unified memory model provides 128 GB of coherent LPDDR5X system memory with up to 273-GB/s bandwidth and 16 memory channels. Meanwhile, the Blackwell GPU includes fifth-generation Tensor Cores with FP4 support for up to 1,000 trillion operations per second of inference throughput.

The CPU subsystem uses a hybrid Arm architecture combining ten Cortex-X925 performance cores with ten Cortex-A725 efficiency cores. The hybrid configuration is one of the features that makes the system viable for desktops, since it optimizes both compute and thermal performance within a 240-W power envelope. The system also includes 4 TB of encrypted NVMe storage, Wi-Fi 7, Bluetooth 5.4, and high-speed USB-C interfaces.

Nvidia DGX Spark Software Stack

The DGX Spark's software environment is very similar to the industrial-scale Nvidia AI platform. The system is built around the Nvidia DGX OS, a Linux-based operating system optimized for AI workloads and preloaded with components of the Nvidia AI Enterprise suite, such as the CUDA platform, cuDNN for deep learning primitives, and TensorRT for optimized inference execution. The software stack also provides access to high-throughput matrix and collective communications libraries, such as cuBLAS and NCCL.

The DGX Spark software stack. 

Developers can leverage pre-installed support for PyTorch, TensorFlow, and MATLAB. The system also supports containerized environments through Docker and the Nvidia Container Toolkit. The system is also integrated with AI Workbench and NGC for model versioning, deployment templates, and access to pretrained models and scripts.

Another unique feature of the software stack is that DGX Spark supports Nvidia NIM, a suite of AI microservices that simplify the deployment of language, vision, and multimodal AI agents. Combined with tools like Ollama and frameworks including Riva and Metropolis, developers can customize inference engines and user-facing applications directly on their local systems. According to Nvidia, such integration simplifies scalability from desktop development to production deployment in DGX Cloud.

Compact Power, Broad Implications

DGX Spark takes an interesting direction in how AI development may be distributed across industries. With petascale performance and a small form factor, it can enable researchers and developers to work locally with models previously constrained to data centers. In that way, DGX Spark has the potential to democratize computing in fields like academic research and scientific experimentation. 

DGX Spark is now available through Nvidia and partner vendors, including Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, MSI, and Micro Center in the U.S. Nvidia's global partners also support global availability.

All images used courtesy of Nvidia.