Nvidia reveals DGX Spark: The world’s smallest AI supercomputer

Nvidia unveiled its DGX Spark and DGX Station personal AI supercomputers at the GTC conference, designed for handling large AI models with or without a datacenter connection. The DGX Spark is now available for preorder.

The DGX Spark, priced at $3,000, is touted as the “world’s smallest AI supercomputer,” previously known as “Digits,” which was announced at CES earlier this year. The larger DGX Station, which has no current price tag, targets AI developers, researchers, data scientists, and students for prototyping, fine-tuning, and inference of large models on desktops.

Powering the DGX Spark is Nvidia’s GB10 Blackwell Superchip, equipped with a GPU that includes fifth-generation Tensor Cores and FP4 support. The GB10 is specifically optimized for the Spark’s compact form factor and can achieve up to 1,000 trillion operations per second (TOPS) for tasks like fine-tuning and inference on advanced AI reasoning models, including the NVIDIA Cosmos Reason world foundation model and the NVIDIA GR00T N1 robot foundation model. The DGX Spark features 128GB of unified memory and provides up to 4TB of NVMe SSD storage.

Nvidia can boost DeepSeek R1’s speed 30x, says Jensen Huang

Conversely, the larger DGX Station utilizes the newly announced GB300 Blackwell Ultra desktop super chip, offering 20 petaflops of AI performance alongside 784GB of unified system memory.

Nvidia also announced that Original Equipment Manufacturer (OEM) partners, including Asus, Dell, HP, Boxx, Lambda, and Supermicro, will produce versions of the DGX computers. The DGX Station will be available from these partners later this year, while versions of the DGX Spark will be offered by Asus, Dell, HP, and Lenovo. Preorders for the DGX Spark can be made on Nvidia’s website, with anticipated deliveries in the summer.

Other companies are also developing GPUs with significant unified memory for local large language models (LLMs). AMD has the Ryzen AI Max+ “Strix Halo,” while HP has incorporated the 128GB version into a laptop, and Framework has included it in a $2,000 desktop, allowing the GPU to access up to 96GB of VRAM.

Featured image credit: Nvidia