CES 2025: Project DIGITS NVIDIA presents AI desktop "supercomputer"

Related Vendors

NVIDIA GmbH

Wuxi InfiMotion Technology Co., Ltd.

NVIDIA has unveiled its first own desktop computer at CES 2025 in Las Vegas: The NVIDIA DIGITS computer, which is based on the in-house GB10 processor, promises 1 petaflop of AI performance with FP4 precision and is said to be comparable in price to a high-end gaming PC.

Thanks to the AI boom, NVIDIA has become the most valuable chip manufacturer in the world thanks to its processors optimized for accelerating artificial intelligence. Now the GPU specialist wants to crown its dominance in the AI hardware sector. At CES 2025 in Las Vegas, the company presented its first desktop PC that has been optimized for processing and accelerating extensive AI models with up to 200 billion parameters (so-called 200B models).

20 ARM cores, 128 GByte DDR5 memory, up to 1 petaflop computing power

The "Project DIGITS" computer, described by NVIDIA as a desktop AI "supercomputer", is based on the company's own newly developed GB10 processor from the "Grace Blackwell" architecture. This is a combination of the proven Blackwell GPU series in the AI sector with an ARM-based "Grace" CPU with 20 processor cores in one housing, which are connected to each other via an integrated NVLink interconnect to ensure the highest possible data transfer rates.

In addition, the computer has 128 GB of DDR5 memory ("DDR5 Unified Memory") to support the fast processing of AI workloads. NVIDIA states that the DIGITS desktop computer can process 200B AI models in 4-bit INT4 format with a performance of up to 1 petaflops. Support for up to 4 TByte of local NVMe storage is also planned. In addition, two DIGITS desktop computers can be networked together via a ConnectX link, an in-house network protocol from NVIDIA. The computer also has USB interfaces and WiFi connectivity.

Develop 200B AI models locally and scale them in the cloud or data centers

The DIGITS computer is designed to enable companies and researchers to prototype, fine-tune and test their own complex models on local systems using the NVIDIA DGX operating system. The NVIDIA DGX OS is an offshoot of the Linux distribution Ubuntu. These models can then be seamlessly deployed on accelerated cloud instances or data center infrastructures. Direct integration into the in-house NVIDIA DGX Cloud is also planned. Developers would thus be able to create their own AI prototypes on Project DIGITS in a relatively straightforward way and then scale them to cloud or data center infrastructures using the same Grace Blackwell architecture and NVIDIA AI Enterprise software platform. Inference or the execution of local models in applications such as robotics and computer vision is also possible.

Project DIGITS users can access an extensive AI library from NVIDIA for experimentation and prototyping. This includes software development kits, frameworks and models available in the NVIDIA NGC catalog and on the NVIDIA Developer Portal. Developers can further refine their own models with the NVIDIA NeMo framework, accelerate data science with NVIDIA RAPIDS libraries and use common frameworks such as PyTorch, Python and Jupyter notebooks.

To develop agent-based AI applications, users can also leverage the company's blueprint templates and NVIDIA's NIM microservices, which are available for research, development and testing through the NVIDIA Developer Program. For enterprise-level professional use, NVIDIA offers an AI Enterprise license for additional security, support and future product releases of NVIDIA AI software.

Project DIGITS is scheduled to go on sale in May 2025 and will be available directly from NVIDIA and licensed partners. According to NVIDIA, the expected retail price is USD 3,000 for the basic model.(sg)