NVIDIA has begun shipping the DGX Spark, which is the compact AI supercomputer designed to bring petaflop-level performance to the desktop.
The system aims to empower developers with the capability to train and fine-tune large AI models locally, without depending on cloud infrastructure or large-scale data centers. AS such, its GB10 Superchip is capable of pushing 1 petaflop of AI performance as well as hosting 128GB of unified memory that handles 200B parameter inferencing, or 70B parameter fine-tuning.
Additionally, through NVIDIA ConnectX-7 200 Gb/s networking and NVLink-C2C technology, multiple DGX Sparks can come together and offer more performance without hassle, and of course, NVIDIA’s entire AI software stack can be accessed in equal terms just like the server-class products.
And that means things like CUDA libraries, CUDA libraries, NVIDIA NIM microservices, and models such as FLUX.1 from Black Forest Labs and Cosmos Reason are all accessible to make development more affordable than ever.
To commemorate the launch, NVIDIA CEO Jensen Huang hand-delivered one of the first DGX Spark systems to Elon Musk at SpaceX’s Starbase in Texas, echoing the symbolic gesture from 2016 that marked the beginning of NVIDIA’s DGX lineage. The company says DGX Spark is part of its broader mission to make advanced AI computing accessible to developers worldwide.
