Nvidia has unveiled its GDX series “personal AI supercomputer” and workstation, both built on Grace Blackwell superchips. Nvidia’s Blackwell Ultra-based desktop supercomputers are aimed at AI developers, researchers, and others who prototype, tune, and refine large AI models.
Image source: NVIDIA
DGX Spark (formerly Project DIGITS) and DGX Station bring the power of the Grace Blackwell architecture, previously only available in data centers, to the desktop. They come with a full suite of software designed to create, tune, and infer language models. The DGX series is being developed in collaboration with Nvidia partners including Asus, Dell, HP, and Lenovo.
«“AI has transformed every layer of the computing stack. It makes sense that a new class of computers will emerge that are built for AI developers and can also be used to run AI applications. With the new DGX personal AI computers, AI can extend from the cloud to desktop and edge applications,” said Jensen Huang, founder and CEO of Nvidia.
DGX Spark is the world’s smallest AI supercomputer, based on the GB10 superchip, optimized for a desktop form factor. The chip is equipped with a powerful graphics processor with fifth-generation Tensor Cores, it provides FP4 performance of up to 1 petaflops (1000 trillion operations per second). The GB10 superchip includes a Blackwell accelerator, which is connected via NVLink-C2C to the processor, unified coherent memory and storage, due to which the chip has access to data between the GPU and CPU to optimize performance under heavy loads using large amounts of memory.
Nvidia’s full-stack AI platform will allow DGX Spark users to seamlessly migrate their AI models from their workstations to DGX Cloud or any other cloud or data center infrastructure with virtually no code changes, making it easier to prototype, tune, and iterate on workflows.
The DGX Station delivers data center-class performance in a compact AI computer. It is the first workstation based on the GB300 accelerator based on the Grace Blackwell architecture. The Blackwell Ultra GPU with the latest generation of Tensor Cores is connected to the processor via NVLink-C2C, providing best-in-class performance.
DGX Station features the ConnectX-8 SuperNIC, which is optimized to accelerate hyperscale compute workloads with support for network speeds up to 800 Gbps. By combining the capabilities of DGX workstations with the NVIDIA CUDA-X AI platform, developers can achieve high levels of performance when building AI models on the PC.