NVIDIA unveiled Project Digits, its latest and smallest AI supercomputer, at the Consumer Electronics Show (CES) 2025. This device can be connected wired or wirelessly and execute AI tasks locally, just like in the cloud.

It also supports popular programming frameworks such as PyTorch, Python, and Jupyter Notebooks. Furthermore, it includes the NVIDIA AI platform, which offers development tools, pre-trained models, and services available on its cloud platform.

Jensen Huang stated that this device is suitable for every computer user. AI has become an integral part of everyone’s work, not just programmers and engineers who develop, train, and test AI models, or content creators and artists who use it to create visual content. It is now embedded in many services that any user needs.

Digits will be released in May (under a new name) at a price of $3,000. It boasts a computing power of 1 petaflop (FP4 floating-point precision). This enables it to handle large AI models with up to 200 billion parameters. Two Digits devices can be linked together to work as a single unit, handling even larger models with 405 billion parameters, such as Meta’s latest open-source Llama 3.1 model.

The device is powered by a new processor, a chip that combines two chips: the first is the Grace CPU developed by MediaTek, and the second is the Blackwell GPU, forming the GB10 Grace Blackwell superchip. It features 128 GB of RAM and a storage drive with a capacity of up to 4 terabytes.

It’s worth noting that the power of this small device is equivalent to that of NVIDIA’s first AI supercomputer, DGX-1, which was delivered to OpenAI in 2016. It was a large device weighing approximately 70 kilograms. This is a clear visual indicator of the significant progress NVIDIA has made in developing the capabilities of AI processors in recent years.