NVIDIA's Project DIGITS Makes Grace Blackwell Supercomputer Available to AI Devs

NVIDIA has a history of making some small, power-efficient development hardware for AI developers, data scientists, and students. But get ready for something much bigger than past efforts. At CES 2025 during tonight's keynote, NVIDIA CEO Jensen Huang announced a huge expansion to the company's previous DIGITS DevBox efforts. Dubbed Project DIGITS, the company released an Arm64-based PC with NVIDIA's own GPU technology to take the next leap in AI development kit technology. 

Long-time HotHardware readers are no doubt familiar with the Jetson AGX Orin Development Kit, which brought some serious Ampere horsepower in a small, power-efficient package. Project DIGITS is kind of like that, except that it's giving developers access to a petaflop of compute power, with the caveat that we're talking about four-bit integers here (INT4), not larger floating point numbers. NVIDIA calls it a supercomputer, and while it's a lot smaller than most supercomputers, it's likely going to be pretty powerful for its pint size. 


By contrast, NVIDIA specs the AGX Orin as being capable of 5.3 TFLOPs of FLOAT32. Different kinds of math are not comparable, but we can assume that Project DIGITS' technology is much faster than the older Ampere architecture (and of course it is since it skips a generation). Still, it's got to have a pretty small GPU in a relatively modest power envelope.

Speaking of architectures, NVIDIA is calling the architecture of the GB10 Superchip is an SoC Grace Blackwell. As the name implies, this means a Blackwell GPU much like the GeForce 50 series. That brings with it the company's latest-generation CUDA cores and brand new fifth-generation Tensor cores are along for the ride. The GPU is connected via an NVLink chip-to-chip interconnect to the Grace CPU, which is a 20-core Arm design that NVIDIA worked on in collaboration with MediaTek. 


AI is a large task that requires gobs of memory, and each Project DIGITS kit includes 128 GB of unified, coherent memory and a maximum of 4 TB of NVMe storage. If that's not enough storage and compute power, a pair of Project DIGITS can be networked via NVIDIA's ConnectX to combine their powers and run models with up to 405 billion parameters. We have a feeling this machine (or perhaps a pair of them) will work well with NVIDIA's Cosmos world foundation models for physical generative AI. 

Project DIGITS will ship running NVIDIA's Linux-based DGX OS. The point of these systems is to prototype heavy-duty AI on a Project DIGITS box and then scale it using some form of distributed compute platform. Models developed on these systems are then seamlessly deployed to NVIDIA's DGX Cloud or other accelerated cloud instances, or to private datacenter infrastructure. 

NVIDIA says that Project DIGITS machines will start with a $3,000 price tag and will be available starting in May from the company itself or its partners. You can head on over to the Project DIGITS website and register for availability notifications now.