Nvidia announces the availability of the DGX Spark: a 1.2-kilogram minicomputer delivering a petaflop of AI computing power.
Nvidia officially puts the DGX Spark on store shelves. Jensen Huang personally delivered one of the first models to Elon Musk and SpaceX. With the DGX Spark, Nvidia aims to put supercomputer power in the palm of your hand. The mini workstation is equipped with a custom Grace Blackwell chip to deliver up to 1 petaflop of local AI computing power.
The DGX Spark is about the size of a folded origami paper and as thick as a book cover, Nvidia writes. But don’t underestimate this small 1.2-kilogram workstation. At the computer’s core sits a ‘shrunk’ GB10 chip. The chip consists of two dies connected via NVLink. Nvidia developed the chip in collaboration with MediaTek.
1 Petaflop
The GB10 can deliver up to 1 petaflop of computing power for AI workloads, sufficient to process models up to 200 billion parameters. The CPU and GPU are powered by shared LPDDR5x memory for local fine-tuning and inference. Those fortunate enough to get their hands on two DGX Spark computers can connect them using the integrated ConnectX network card to double the capabilities.
The Nvidia DGX Spark will cost approximately $3,000. The workstation isn’t aimed at the mass market, but rather at data scientists, researchers, and robotics engineers seeking a cost-effective solution for running heavy AI workloads. Nvidia’s most advanced graphics cards already cost twice as much.
Loyal Nvidia partners such as Lenovo, HP, Dell, and Gigabyte will launch their own Spark systems later this week. Nvidia has an even more powerful version, the DGX Station, in the pipeline. It will be equipped with a GB300 chip capable of delivering up to 20 petaflops of AI computing power.