Nvidia introduces Grace-Blackwell variant, an optimized superchip for data center AI workloads, expanding the Blackwell product family.

Nvidia has announced Project Digits, a $3,000 desktop system powered by a new GB10 Grace-Blackwell superchip and equipped with 128GB of memory, designed to give AI developers, researchers, and students the tools to run large models on the desktop. Developed in collaboration with MediaTek and unveiled at CES in Las Vegas by CEO Jensen Huang, the system features an Arm-based Grace CPU and Blackwell GPU that appear to reside in a single SoC. The box will ship with a special build of Ubuntu Linux pre-configured to take advantage of the hardware. Resembling an Intel NUC mini-PC in size, Project Digits delivers a full petaFLOP of AI performance, though this measurement is based on sparse 4-bit floating point workloads.

Specifications suggest the GB10 features a 20-core Grace CPU and a GPU that manages 40th the performance of the twin Blackwell GPUs used in Nvidia's GB200 AI server. The system packs more power than AI PCs from Intel, AMD, or Qualcomm, but will struggle to compete with workstations equipped with Nvidia's RTX 6000 Ada, which boasts 1.45 petaFLOPS of sparse FP/INT8 performance—roughly triple the estimated 500 teraFLOPS Project Digits will deliver at that precision. The 128GB of LPDDR5x memory was intentionally chosen to make working with large AI models more accessible, according to Allen Bourgoyne, Nvidia's director of product marketing for enterprise platforms.

Nvidia claims Project Digits will support models up to 200 billion parameters, provided they are compressed to 4-bits. Two units can be connected via onboard ConnectX networking to run models with up to 405 billion parameters, putting Meta's Llama 3.1 405B within reach—again at 4-bits. For reference, running that same model on existing workstation hardware would require at least five 48GB GPUs. Based on renders showing six LPDDR5x modules and assuming memory speeds of 8,800 MT/s, the system would deliver approximately 825GB/s of bandwidth, not far off from the RTX 6000 Ada's 960GB/s, which could translate to around eight tokens per second for a 200 billion parameter model—though full specifications were not available at the time of announcement.

Beyond AI inference, Nvidia expects Project Digits will suit model experimentation, fine-tuning, data science, and edge applications. The system will include 4TB of NVMe storage, ample capacity for most open models, especially those quantized to 4-bits. Project Digits is available beginning in May and represents a grown-up version of Nvidia's Jetson dev kits, which the company has offered for years and recently refreshed with the Orin Nano Super in December 2024. In its current form, the machine appears aimed at familiarizing users with Nvidia's more powerful Grace-Blackwell superchips like the GB200 and GB200 NVL4, though Nvidia has not disclosed whether it will license the GB10 to other PC makers.