Learn · NVIDIA Hardware & DGX · The DGX Family: Spark, Station
DGX Station: The Trillion-Parameter Desk Beast
In plain English
If Spark is a sprinter, DGX Station is a whale with a PhD. Its GB300 chip has a GPU so big it gets 252GB of super-fast memory of its own, plus a 496GB CPU memory pool next door — 748GB total. That's enough to hold an AI with 1 TRILLION parameters on someone's desk. 💪
What is actually going on
DGX Station runs the GB300 Grace Blackwell Ultra Desktop Superchip: a Blackwell Ultra GPU (252GB HBM3e at ~7.1TB/s) fused to a 72-core Grace CPU (496GB LPDDR5X) via 900GB/s NVLink-C2C — ~748GB coherent memory. Rated ~20 PFLOPS NVFP4 (FP4 sparse), it serves models up to ~1T parameters, MIG-partitions into 7 instances for teams, and links two Stations over ConnectX-8 (800Gb/s). Built by OEM partners (Asus, Dell, HP, Lenovo, MSI, Pegatron); Windows edition available.
Why it matters
DGX Station is the ceiling of local AI: frontier open-weight models fully local, always-on agents, private enterprise inference — no token bills, no data leaving the building. It's the reference point for what 'local' now means.
Another way to think about it
A private moon. It's big, it's expensive, and everything about it is 'more' — but you can walk to work every day without leaving your desk. 🌙
Check yourself
Which superchip powers NVIDIA DGX Station?
- GB10 Grace Blackwell
- GB300 Grace Blackwell Ultra
- Apple M3 Ultra
- RTX 4090
Show answer
GB300 Grace Blackwell Ultra — DGX Station is the first system on GB300: Blackwell Ultra GPU + 72-core Grace CPU via NVLink-C2C.
Roughly how large a model can DGX Station hold locally?
- 7B parameters
- 70B parameters
- 405B parameters
- 1 trillion parameters
Show answer
1 trillion parameters — With ~748GB coherent memory and NVFP4 quantization, DGX Station supports models up to ~1T parameters.
DGX Station's 748GB total memory means the GPU's 252GB HBM and the CPU's LPDDR5X are the same physical pool with identical speed.
- True
- False
Show answer
False — It's coherent but tiered: 252GB HBM3e at ~7.1TB/s is blazing; the CPU's LPDDR5X is slower to reach over NVLink-C2C — more like unified memory with two speeds.