NVIDIA Puts "a Petaflop of AI Computing Performance" on Anyone's Desk with Project DIGITS

NVIDIA has taken to the stage at the Consumer Electronics Show (CES) in Las Vegas to announce new products and features for those interested in artificial intelligence (AI) — including Project DIGITS, a compact computer that the company claims puts "a petaflop of AI computing performance" on any desk.

"AI will be mainstream in every application for every industry. With Project DIGITS, the Grace Blackwell Superchip comes to millions of developers," claims NVIDIA founder and chief executive officer Jensen Huang of the company's latest launch, powered by its latest graphics processor. "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."

Project DIGITS is, at first glance, a fairly standard compact computer — similar in size to those based around Intel's Next Unit of Computing (NUC) standard. Its internals, though, couldn't be different: the device is powered by NVIDIA's "GB10 Superchip," a system-on-chip based on the company's Blackwell graphics architecture and fifth-generation Tensor Core design plus 20 NVIDIA Grace Arm CPU cores — delivering, all told, a claimed petaflop of compute at FP4 precision.

Coupled with 128GB of unified High-Bandwidth Memory (HBM) and up to 4TB of Non-Volatile Memory Express (NVMe) storage, a Project DIGITS box is claimed to be able to run large language models (LLMs) on-device at sizes up to 200 billion parameters. For those looking to go even larger, the company says that two machines can be linked over ConnectX networking in order to share resources and run LLMs of up to 405 billion parameters in size entirely on-device.

Other announcements made during the company's AI-heavy keynote presentation include the launch of Blackwell-based desktop and laptop GeForce RTX 5000-series graphics processors, the first to support the same FP4 precision as the Project DIGITS system-on-chip. The top-end model, the GeForce RTX 5090, will deliver a claimed 3,352 tera-operations per second (TOPS) of compute, with the RTX 5080 offering 1,801 TOPS, the RTX 5070 Ti at 1,406 TOPS, and the RTX 5070 at 988 TOPS.

The company also unveiled foundation models, running as NIM microservices, targeting the new RTX 5000-series GPUs and including large language models, vision language models, image generators, computer vision, and more. The company also unveiled Llama Nemotron, a model targeting "agentic tasks" — where the model itself is expected to carry out the steps of a procedure, rather than simply describing it to the user. The Cosmos foundation model, meanwhile, is claimed to be trained on over 20 million hours of video footage and aims to speed the development of "physical AI" — where the AI is directly powering robots and autonomous vehicles.

"The ChatGPT moment for robotics is coming. Like large language models, world foundation models are fundamental to advancing robot and AV development, yet not all developers have the expertise and resources to train their own," claims Huang. "We created Cosmos to democratize physical AI and put general robotics in reach of every developer."

The GeForce RTX 5090 and RTX 5080 GPUs will be available from 30 January, priced at $1,999 and $999 respectively, NVIDIA has confirmed; the RTX 5070 Ti and RTX 5070 will follow in February at $749 and $549. Project DIGITS systems will be sold directly by NVIDIA and "top partners" starting in May, the company says, with pricing beginning at $3,000 per unit. Interested parties can sign up to be notified when the device goes up for sale on the NVIDIA website.