NVIDIA introduced a personal supercomputer for artificial intelligence – Project Digits. The announcement took place at the CES exhibition, and the start of sales is scheduled for May. The main element of the system is the new GB10 Grace Blackwell Superchip, which has a power sufficient to work with modern AI models. The device is compactly, takes a place on the table and works on a standard outlet, replacing bulky energy-intensive servers of the past. The system supports models up to 200 billion parameters and will cost from $3.000. By design, Project Digits resembles Mac mini.
Each supercomputer is equipped with 128 GB of RAM and up to 4 TB NVMe drives. For especially complex tasks, two devices can be combined to process models with parameters up to 405 billion. The GB10 processor is capable of reaching 1 petaflop by FP4-Presion calculations, the use of CUDA and Tensor-Judasur of the latest generation and energy-saving ARM-Yadraram developed in partnership with MediaTek.
The system operates on Linux with NVIDIA DGX OS and supports popular frameworks, such as Pytorch, Python, and Jupyter. The kit includes instruments for orchestration and pre-trained models from NVIDIA NGC. For local development and testing of models, the NVIDIA NEMO platform is used, and the acceleration of scientific calculations is provided by the libraries of NVIDIA Rapids.
The project is aimed at developers, researchers, and students, providing them with access to powerful tools for working with AI. As CEO Nvidia Jensen Juang stated: “Project Digits is a way to give every developer, researcher, and student the opportunity to form an era of artificial intelligence.”
NVIDIA also reminded of other available products such as Jetson Orin Nano Super for $249, designed for startups and amateurs capable of processing models up to 8 billion parameters.