It can run 200bn-parameter models which can be seamlessl
y deployed on accelerated cloud or data centre infrastructure. This increases to up to 405bn parameter models when using ConnectX networking to link two computers.
The superchip has an AI computing performance of up to 1pflop at FP4 precision. Its Blackwell GPU, with CUDA cores and fifth generation Tensor cores, are connected to a Grace CPU via NVLink-C2 chip-to-chip interconnect.
The GB10 performance means that a supercomputer can use a standard electrical outlet and systems running Linux-based Nvidia DGX OS for deployment on DGX Cloud. The Project Digits supercomputer has 128GB memory and up to 4TB of NVMe storage. “Placing an AI supercomputer on the desks of every data scientist, AI researcher and student empowers them to engage and shape the age of AI,” said Jensen Huang, founder and CEO of Nvidia at the launch of Project Digits at CES in Las Vegas.
AI models can be fine-tuned with the company’s NeMo framework, Rapids libraries and running frameworks such as Pytorch, Python and Jupyter notebooks. Nvidia also announced new Blueprints and NIM microservices for agentic AI (i.e. AI decision making with minimal direct human supervision).
Project Digits machines are scheduled for availability from May 2025.
