NVIDIA has taken to the stage on the Shopper Electronics Present (CES) in Las Vegas to announce new merchandise and options for these excited by synthetic intelligence (AI) — together with Challenge DIGITS, a compact pc that the corporate claims places “a petaflop of AI computing efficiency” on any desk.
“AI shall be mainstream in each utility for each trade. With Challenge DIGITS, the Grace Blackwell Superchip involves thousands and thousands of builders,” claims NVIDIA founder and chief government officer Jensen Huang of the corporate’s newest launch, powered by its newest graphics processor. “Inserting an AI supercomputer on the desks of each knowledge scientist, AI researcher, and scholar empowers them to have interaction and form the age of AI.”
NVIDIA’s Challenge DIGITS is a supercomputer in miniature, delivering a claimed petaflop of FP4 compute in a NUC-like housing. (📷: NVIDIA)
Challenge DIGITS is, at first look, a reasonably commonplace compact pc — related in dimension to these based mostly round Intel’s Subsequent Unit of Computing (NUC) commonplace. Its internals, although, could not be totally different: the system is powered by NVIDIA’s “GB10 Superchip,” a system-on-chip based mostly on the corporate’s Blackwell graphics structure and fifth-generation Tensor Core design plus 20 NVIDIA Grace Arm CPU cores — delivering, all instructed, a claimed petaflop of compute at FP4 precision.
Coupled with 128GB of unified Excessive-Bandwidth Reminiscence (HBM) and as much as 4TB of Non-Unstable Reminiscence Categorical (NVMe) storage, a Challenge DIGITS field is claimed to have the ability to run giant language fashions (LLMs) on-device at sizes as much as 200 billion parameters. For these seeking to go even bigger, the corporate says that two machines will be linked over ConnectX networking with a purpose to share sources and run LLMs of as much as 405 billion parameters in dimension totally on-device.
The corporate’s CES keynote additionally noticed the revealing of the RTX 5000-series of GPUs, that includes as much as 3,352 TOPS of AI compute. (📷: NVIDIA)
Different bulletins made throughout the firm’s AI-heavy keynote presentation embody the launch of Blackwell-based desktop and laptop computer GeForce RTX 5000-series graphics processors, the primary to help the identical FP4 precision because the Challenge DIGITS system-on-chip. The highest-end mannequin, the GeForce RTX 5090, will ship a claimed 3,352 tera-operations per second (TOPS) of compute, with the RTX 5080 providing 1,801 TOPS, the RTX 5070 Ti at 1,406 TOPS, and the RTX 5070 at 988 TOPS.
The corporate additionally unveiled basis fashions, operating as NIM microservices, concentrating on the brand new RTX 5000-series GPUs and together with giant language fashions, imaginative and prescient language fashions, picture turbines, pc imaginative and prescient, and extra. The corporate additionally unveiled Llama Nemotron, a mannequin concentrating on “agentic duties” — the place the mannequin itself is anticipated to hold out the steps of a process, reasonably than merely describing it to the person. The Cosmos basis mannequin, in the meantime, is claimed to be skilled on over 20 million hours of video footage and goals to hurry the event of “bodily AI” — the place the AI is immediately powering robots and autonomous autos.
Different bulletins included the revealing of basis fashions, together with Cosmos (pictured). (📷: NVIDIA)
“The ChatGPT second for robotics is coming. Like giant language fashions, world basis fashions are elementary to advancing robotic and AV growth, but not all builders have the experience and sources to coach their very own,” claims Huang. “We created Cosmos to democratize bodily AI and put normal robotics in attain of each developer.”
The GeForce RTX 5090 and RTX 5080 GPUs shall be accessible from 30 January, priced at $1,999 and $999 respectively, NVIDIA has confirmed; the RTX 5070 Ti and RTX 5070 will comply with in February at $749 and $549. Challenge DIGITS techniques shall be offered immediately by NVIDIA and “prime companions” beginning in Could, the corporate says, with pricing starting at $3,000 per unit. events can signal as much as be notified when the system goes up on the market on the NVIDIA web site.
