NVIDIA DGX Spark - a revolution in access to AI supercomputing power

The launch of the DGX Spark marks the introduction of a new category of computers aimed at developers of solutions based on artificial intelligence. The system offers a petaflop of computing power (one trillion floating point operations performed per second) and 128 GB of unified system memory (shared by the CPU and GPU) in a chassis compact enough to fit on a standard desktop. The manufacturer describes it as the smallest AI supercomputer available on the market.

The device is designed for programmers, researchers and developers who need supercomputer-class performance. Using NVIDIA's Grace Blackwell architecture, the system integrates GPUs, CPUs, networking technologies, CUDA libraries and AI software into a single, compact solution.

At the heart of the DGX Spark system is the NVIDIA GB10 Grace Blackwell superchip, an innovative system-on-chip (SoC) that combines a CPU and graphics card on a single chip. The architecture uses a 20-core ARM processor in a big-little configuration with 10 Cortex-X925 cores and 10 Cortex-A725 cores, while offering 128GB of memory shared by the CPU and GPU.

The graphics part, based on the Blackwell architecture, contains 6144 CUDA cores, fifth-generation Tensor cores (with FP4 support) and fourth-generation RT cores. The system delivers impressive computational performance of 1,000 TOPS (trillion operations per second) and 1 PFLOP (trillion operations per second) at FP4 precision.

The DGX Spark features 128 GB of unified LPDDR5x memory, running at 4266 MHz on a 256-bit bus, providing 273 GB/s throughput. This unified memory architecture allows the CPU and GPU to share the same memory area thanks to NVLink-C2C technology, which offers five times the bandwidth of PCIe 5.0.

The data storage system is based on an NVMe M.2 drive, with 4TB of data space and support for hardware encryption, providing enough space to store large AI models and the data sets needed for training and inference (the process of generating responses or inferences based on input data).

The computational capabilities of the device allow AI models to work locally with up to 200 billion parameters in inference mode and up to 70 billion parameters in the tuning process. The system also makes it possible to create AI agents and run advanced software stacks without the need for cloud computing or local data centers.

One of the key advantages of the DGX Spark is its pre-installed full NVIDIA AI software stack.

This includes CUDA libraries or models and NVIDIA NIM microservices, enabling them to start working on AI projects immediately. Developers also get access to NVIDIA AI ecosystem tools, including Black Forest Labs FLUX.1 models for image generation, NVIDIA Cosmos Reason for visual analysis, and the optimized Qwen3 chatbot.

A day before the DGX Spark began shipping worldwide, Jensen Huang, founder and CEO of NVIDIA, personally delivered one of the first units of the device to Elon Musk at Starbase in Texas. The event was a reference to the origins of DGX supercomputers - in 2016, Huang also personally handed over the first NVIDIA DGX-1 system to Musk, who at the time was running OpenAI.

For the sake of comparison, let's take a peek here at the gap between these devices in terms of performance and capabilities over this less than a decade:

As Huang emphasized, DGX-1 ushered in the era of AI supercomputers that drive modern artificial intelligence. From the first system delivered to OpenAI, ChatGPT was born, which is what started the AI revolution. DGX Spark is set to follow up on that mission - putting an AI computer in the hands of every developer to start the next wave of breakthroughs.