NVIDIA GB10 Grace Blackwell Superchip

NVIDIA has unveiled the GB10 Grace Blackwell Superchip, a groundbreaking system-on-a-chip (SoC) that delivers up to 1 petaflop of AI performance at FP4 precision. This innovation is central to Project DIGITS, NVIDIA’s personal AI supercomputer designed for AI researchers, data scientists, and students.

Key Features of the GB10 Superchip:

• Architecture: The GB10 integrates an NVIDIA Blackwell GPU with the latest CUDA cores and fifth-generation Tensor Cores, connected via NVLink-C2C to a high-performance NVIDIA Grace CPU comprising 20 Arm-based cores. This design, developed in collaboration with MediaTek, ensures exceptional power efficiency and performance.

• Memory and Storage: Each Project DIGITS system equipped with the GB10 Superchip offers 128GB of unified, coherent memory and up to 4TB of NVMe storage, enabling the handling of large AI models with up to 200 billion parameters.

• Scalability: For more demanding applications, two Project DIGITS systems can be interconnected using NVIDIA ConnectX networking to support models with up to 405 billion parameters.

• Software Integration: The system runs on a Linux-based NVIDIA DGX OS and provides access to NVIDIA’s extensive AI software library, including frameworks like PyTorch, Python, and Jupyter notebooks. This seamless integration allows developers to prototype AI on Project DIGITS and scale to cloud or data center infrastructure using the same architecture and software platform.

Availability and Pricing:

Project DIGITS, powered by the GB10 Superchip, is scheduled for release in May 2025, with prices starting at $3,000. This initiative aims to democratize AI capabilities, making advanced AI development accessible to a broader range of users.

NVIDIA’s introduction of the GB10 Grace Blackwell Superchip marks a significant advancement in AI computing, providing unprecedented performance and accessibility for developers and researchers worldwide.