AI Supercomputer from NVIDIA that can run 200B-Parameter Models

NVIDIA presented the smallest supercomputer for AI tasks. Project Digits is mainly oriented on data scientists, AI researchers, and just users around the globe who would like to test Grace Blackwell.  

GB10 Grace Blackwell Superchips

The peculiarity of the project is in GB10 Grace Blackwell Superchip which has petaflot performance for working with huge AI models, prototyping, and other large tasks.

The GB10 Superchip is based on Grace Blackwell architecture and offers exclusive performance characteristics of 1 petaflop at FP4 precision. The chip’s design is a collaboration of MediaTek and NVIDIA that achieved exceptional results and created a product with exclusive connectivity, power efficiency, and performance. This powerful performance of the Superchip can be achieved by the usage of the standard outlet.

Project DIGITS has 4TB of NVMe storage and 128GB of coherent memory. To achieve higher progress in the innovative approaches, the supercomputer can run nearly 200 billion parameter models. To double the number of running models, it is possible to use 2 linked supercomputers.

Users of Project DIGITS also have access to a huge library of AI software that includes frameworks and various models, orchestration tools, prototyping, and development kits. Among the most common frameworks, are Jupyter and PyTorch.

According to the current info, Project DIGITS will be available at the end of the spring and its price starts at $3,000.