Nvidia’s Project DIGITS puts AI supercomputing chips on the desktop

Nvidia has built its generative AI (genAI) business on delivering massive computing capacity to data centers where it can be used to train and refine large language models (LLMs). 

Now, the company is readying a diminutive desktop device called Project DIGITS, a “personal AI supercomputer” with a lightweight version of the Grace Blackwell platform found in its most powerful servers; it’s aimed at data scientists, researchers, and students who will be able to prototype, tune, and run large genAI models.

Nvidia CEO Jensen Huang unveiled Project DIGITS in a keynote speech on the eve of the CES 2025 electronics show in Las Vegas.

Project DIGITS is similar in size to the Windows 365 Link thin client Microsoft unveiled in November. Microsoft’s Link measures 120mm (4.72 inches) square and is 30mm (1.18 inches) high. 

Nvidia hasn’t given precise dimensions for Project DIGITS, with Allen Bourgoyne, director of product marketing, saying only that the square device will be about as wide as a coffee mug, including the handle, and about half as high. No international standard exists for mugs, but they are typically about 120mm across, including the handle, and around 90mm high, making Project DIGITS as wide as the Link but half as thick again. There the resemblance ends.

The philosophies behind the two devices are quite different: Where the Link pushes almost all the computing capacity to the cloud, Nvidia’s hardware is moving it down to the desktop. 

Microsoft’s Link has 8GB of RAM, no local data storage, and an unspecified Intel processor with no special AI capabilities: If you want to use Windows’ Copilot features they — like everything else — will run in the cloud. Link will sell for around $350 when it goes on sale in April.

One wall outlet, one petaflop

Project DIGITS, on the other hand, will cost upwards of $3,000 when it arrives in May. For that money, buyers will get 4TB of NVMe storage, 128GB of unified Low-Power DDR5X system memory, and a new GB10 Grace Blackwell Superchip; it comes with 20 ARM cores in the Grace CPU and a mix of CUDA cores, RT cores and fifth-generation tensor cores in the Blackwell GPU. 

Together, those cores offer up to 1 petaflop of AI processing capability — enough, said Bourgoyne, to work with a 200-billion-parameter model at “FP4” accuracy locally, with no need for the cloud. By connecting two Project DIGITS devices together via their built-in ConnectX networking chips, it’s possible to work with 400-billion-parameter models, he said.

The GB10 was co-developed with Mediatek, a company known for its power-efficient mobile chips. Compared to the GB200 processors used in data centers, an NV72 rack full of which can draw as much as 120kW, Project DIGITS is more power efficient. “So, you can plug it into a standard wall outlet,” Bourgoyne said. “It doesn’t require any additional power than what you have at your desktop.”

Project DIGITS won’t run Windows: Instead, it will run DGX OS, a version of Ubuntu Linux customized with additional drivers and tools for developing and running AI applications. That’s the same software that runs on Nvidia DGX systems in the data center, meaning models built and tested locally on Project DIGITS can be deployed straight to the cloud or data center, the company said. 

Other Nvidia AI tools the device can run include orchestration tools, frameworks, and models on the Nvidia Developer portal and in its NGC. That includes the NeMo framework for fine-tuning models and the RAPIDS libraries for data science.

Nvidia Blueprints and NIM microservices are available under lightweight licenses via its developer program for building agentic AI applications, with an AI Enterprise license needed only when they are moved to production environments, the company said. 

More generative AI on the desktop

Recognizing that you don’t need a GB10 processor to accelerate AI development on the desktop, Nvidia is also introducing a range of NIM microservices and AI Blueprints for building applications on PCs containing its Geforce RTX GPUs — what it calls RTX AI PCs.

Nvidia is introducing a range of AI foundation models — both its own and those of other developers — containerized as NIM microservices that can be downloaded and connected together. Using low-code and no-code tools such as AnythingLLM, ComfyUI, Langflow and LM Studio, developers will be able to build and deploy workflows using these NIM microservices, with its AI Blueprints providing preconfigured workflows for particular tasks such as converting between media formats. One of the newest Blueprints can convert PDFs to podcasts.