NVIDIA Unveils DGX Station, a Desktop Supercomputer for AI Developers
company NVIDIA announced DGX Station, new Desktop supercomputer for AI developers, which will allow you to launch Large Language Models (LLM) locallyThis device is more powerful than DGX Spark (formerly known as Project DIGITS) and is intended for scientists, researchers and data scientists.
DGX Station is based on GB300 Grace Blackwell Ultra Desktop Superchip, including 72-core Grace processor и GPU Blackwell Ultra, connected through NVLink-C2C с with a throughput of 900 GB/s. This chip supports 496GB LPDDR5X RAM и 288 GB HBM3e GPU-memory, which allows you to run models with 200 billion parameters.
Superchip Blackwell Ultra Provides 1,5x increase in AI FLOPS compared to a regular Blackwell, and FP4 optimization Reduces memory load. Built-in NVMe SSD 1 or 4 TB provides fast access to data, and NVIDIA ConnectX-7 SmartNIC Network Adapter supports the 10GbE Ethernet.
The system operates on NVIDIA DGX OS (modified Ubuntu) and supports Migrating LLM from On-Premise Development to the Cloud. DGX Station will go on sale at the end of 2025 year through third-party manufacturers.
For those looking for a budget solution, RTX 5090 allows you to run models up to 30 billion parameters, but its price exceeds $4000. Even RTX 4060 Ti 16GBsupporting LLM up to 14 billion parametersis worth about $1000.
