Since the beginning of computing history, supercomputers have only been owned by government institutions or private companies that can afford equipment that costs tens of millions. This has made access to training artificial intelligence (AI) models beyond the reach of most developers. Today NVIDIA announced Project DIGITS, which they say is the most affordable AI training supercomputer on the market.
Each Project DIGITS supercomputer is equipped with a GB10 Grace Blackwell processor chip paired with 128GB of memory and up to 4TB of NVMe storage. The GB10 Grace Blackwell chip has 1 petaflop of AI processing power, allowing it to train AI models with 200 billion parameters. If a larger model needs to be trained, two Project DIGITS units can be connected using NVIDIA ConnectX.
The GB10 was developed in collaboration with MediaTek and has 20 Arm-based CPU cores as well as next-generation CUDA cores and fifth-generation Tensor cores.
At first glance, Project DIGITS looks like a mini computer about the size of a Mac Mini. The operating system used is the Linux-based NVIDIA DGX OS. NVIDIA Project DIGITS will be available in May for $3,000 (~RM 13,462).