Desktop DGX station surfaced
GTC highlighted a move to bring serious AI power onto desks — NVIDIA showed a desktop DGX Station boasting 748 GB of coherent memory and up to 20 petaflops of FP4 compute, signaling workstation‑level AI for developers and prosumers. That desktop/workstation push was part of the product mix discussed around the conference and underlines that not all AI compute will live in giant cloud racks (x.com) (cloudnews.tech). Put simply: expect higher‑end developer machines to get much more capable this year.
For years, the rule in artificial intelligence was simple: if your model was huge, your computer had to live in a data center. NVIDIA is now selling a desk-side machine called DGX Station that puts up to 20 petaflops of FP4 artificial intelligence compute and 748 gigabytes of coherent memory next to a monitor instead of in a server rack. (nvidia.com) A model is just a giant pile of numbers, and the hard part is keeping enough of those numbers close enough to the chip that uses them. NVIDIA says DGX Station can support models up to 1 trillion parameters locally, which is why the memory figure matters as much as the compute figure. (nvidia.com) Coherent memory means the central processor and the graphics processor can treat memory more like one shared workbench than two separate closets. NVIDIA ties its Grace central processor and Blackwell Ultra graphics processor together in one GB300 desktop superchip so data does not have to be copied back and forth as often. (nvidia.com 1) (nvidia.com 2) That shared pool is split between 252 gigabytes of high bandwidth memory on the graphics side and 496 gigabytes of LPDDR5X memory on the central processor side. ServeTheHome reported that earlier 2025 plans called for 288 gigabytes of high bandwidth memory, so the shipping 2026 version appears to have slightly less graphics-side memory than the first spec sheet promised. (servethehome.com) (nvidia.com) The “FP4” label is a clue about what this box is built for. Four-bit floating point is a very compressed math format for running and tuning modern artificial intelligence models, so NVIDIA is optimizing this machine for inference and fine-tuning jobs that care about speed and memory footprint more than old-school scientific precision. (nvidia.com) NVIDIA is not pitching this as a one-off science project. Its January 2026 announcement said DGX Station systems would come from ASUS, Boxx, Dell Technologies, GIGABYTE, HP, MSI, and Supermicro in spring 2026, and March coverage showed orders opening through several of those partners. (blogs.nvidia.com) (pcmag.com) There is also a smaller sibling called DGX Spark. NVIDIA says DGX Spark delivers up to 1 petaflop of FP4 performance with 128 gigabytes of memory, while DGX Station jumps to 20 petaflops and 748 gigabytes, which creates a ladder from “serious desktop” to “small office supercomputer.” (nvidia.com 1) (nvidia.com 2) The point of that ladder is to let developers build locally and move up only when they need more scale. NVIDIA’s own pitch is that the same software stack can start on a desk-side system and then move to DGX Cloud or larger data center gear, which is a way of keeping expensive cloud time for the jobs that truly need it. (blogs.nvidia.com) This does not mean giant cloud clusters are going away. It means the line between a workstation and an artificial intelligence server is getting blurry enough that a developer, researcher, or small team can run much larger models in-house than a high-end desktop could handle even a year ago. (nvidia.com) (idc.com)
