Artificial intelligence models are mostly served by data centers consisting of servers with enormous computing power, which can be accessed by customers worldwide based on a cloud topology. The best-known hardware manufacturer in the segment, Nvidia, has now set its sights on bringing AI supercomputers directly to the user’s desktop.
The Santa Clara-based chipmaker announced at CES 2025 its special computer for the above purpose, Project Digits, which uses Nvidia’s brand new GB10 Grace Blackwell Superchip, which includes both an AI processing unit and a CPU. According to the manufacturer, the unit can handle AI models consisting of up to 200 billion parameters locally, while it fits in the palm of a person and works with mains power via a traditional wall connection.
2025: how much money is the home office worth to you?
Here are our 2025 forecasts for the IT labor market.
Starting from May, Project Digits, starting at $3,000, will be available in configurations with 128 GB of fixed, non-expandable system memory and up to 4 TB of NVMe storage in addition to the GB10. The system provides the possibility to connect two units, the resulting computing environment can cope with models with up to 405 billion parameters (this is exactly how many parameters Meta’s currently most advanced Llama 3.1 model has).
According to the manufacturer’s announcement, the GB10, which is considered the brain of the system, provides a petaflops of AI computing capacity with FP4 accuracy – the unit actually consists of a Blackwell GPU with a lower performance than data center solutions, but with the same structure as them, which is the Nvidia NVlink-C2C chip-to- chip technology is connected to a Grace CPU with 20 ARM-based processor cores.
The Project Digits systems based on the Blackwell architecture are of course not home toys, the manufacturer recommends the system mainly to researchers working locally with AI models, students, and companies developing AI models. Thanks to the mini supercomputer, the latter can fine-tune and test their own models locally and then scale them up according to the data center environment.
However, Project Digits is not Nvidia’s first local computer for a similar purpose, as last December the company lifted the veil on Jetson, which also runs AI models locally, albeit with significantly lower performance. Suitable for models with up to eight million parameters, Jetson is accordingly available even for hobbyists – the device is available from $249.
Source: www.hwsw.hu
