Nvidia recently unveiled the DGX Station, a powerful deskside supercomputer capable of running AI models with up to one trillion parameters—similar to the scale of GPT-4—entirely offline. This machine, which boasts 748 gigabytes of coherent memory and 20 petaflops of compute, represents a potential game-changer for startups looking to harness cutting-edge AI without relying on cloud infrastructure.
Announced during the company's GTC conference in San Jose, the DGX Station addresses a growing concern in the AI sector: the need for immense computational power while maintaining control over data and intellectual property. As enterprises increasingly prefer local solutions, Nvidia’s new offering brings high-performance computing directly to individual engineers’ desks.
What 20 petaflops on your desktop actually means
The DGX Station is powered by the new GB300 Grace Blackwell Ultra Desktop Superchip, which integrates a 72-core Grace CPU and a Blackwell Ultra GPU using Nvidia's NVLink-C2C interconnect. This technology enables an impressive 1.8 terabytes per second of coherent bandwidth, seven times faster than PCIe Gen 6, allowing seamless sharing of memory between the CPU and GPU.
With 20 petaflops (20 quadrillion operations per second) of processing capability, the DGX Station would rank among the world’s most powerful machines, traditionally found only in vast data centers. Now, it fits neatly in a workspace.