Nvidia AI powerhouse: See the latest mini AI tech

05.02.2025

3 min read

At CES 2025 in Las Vegas, Nvidia unveiled, among other things, a mini-computer equipped with high-performance components such as the Blackwell-GPU and the GB10-Superchip, which is set to usher in a new era of global AI development.Nvidia has not forgotten the gaming community that helped build the US company into a powerhouse. Accordingly, new graphics cards for gaming PCs and consoles were also presented at this year’s CES. However, in his keynote, CEO and founder Jensen Huang was particularly enthusiastic about artificial intelligence, which has transformed the graphics card manufacturer into one of the world’s largest companies by market capitalization in a short amount of time, alongside autonomous driving.In this context, Huang announced a new era of AI development with the Blackwell-GPU and a small desktop computer reminiscent of the Mac Mini. Software developers, as well as business and private users, are expected to play a significant role. The Blackwell-named Graphics Processing Unit (GPU) is one of the key components of this mini-computer, which has already been dubbed a super-AI-computer and codenamed “Project DIGITS.” As IT-Boltwise reports, Blackwell is expected to surpass the sales of its predecessor, Hopper, early in 2025.

GPU Power of a Supercomputer

At the heart of Project DIGITS lies the GB10 Superchip, an Nvidia proprietary system-on-a-chip (SoC). This chip is designed to work in tandem with the Blackwell GPU and the Nvidia proprietary Grace CPU core to support AI language models with up to 200 billion parameters (or up to 400 billion in a dual setup), as reported by Digital Production.

Nvidia’s mini-computer indeed packs a lot of power at a relatively low purchase price of 3,000 dollars. Features include 128 gigabytes of RAM, an NVMe SSD drive with 4 terabytes of storage, the aforementioned Grace CPU with 20 ARM cores, and the Blackwell GPU.

The latter, with its computing power capable of calculating 1 peta-FLOPS or one quadrillion (10^15) floating-point operations per second, already approaches the capabilities of supercomputers, at least those of older models like the IBM Roadrunner, which set the record in May 2008 with 1.026 peta-FLOPS.

Develop locally, share via GB200

As reported by Golem, the housing and SoC of the mini-computer bear a striking resemblance to Nvidia’s DGX systems for servers and workstations. According to the manufacturer’s vision, developers can use this setup to create AI models locally and then deploy them on GB200 systems for productive use. This is facilitated by a ConnectX network adapter, which enables the connection of two computing clusters.

The GB200 high-performance accelerator from Nvidia, however, has faced criticism until the end of 2024. Equipped KI servers risk overheating, which is surprising given their price tag of 60,000 to 70,000 dollars.

Source image: Adobe Stock / Limitless Visions