System Overview#

The NVIDIA DGX Station is an ARM-based workstation built around the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. NVIDIA NVLink-C2C connects the NVIDIA Grace CPU and Blackwell Ultra GPU.

Its purpose is to provide AI developers, data scientists and researches with a desktop-sized device that features the same architecture found in cloud-scale NVIDIA solutions. Applications can be developed and debugged locally with DGX Station and then brought to massive scale on the cloud. Both desktop and server use cases are supported.

Hardware#

DGX Station is based on the Grace Blackwell architecture capable of up to 20 PFLOPs of AI compute. The system contains up to 748 GB of coherent memory and enables work with huge AI models of up to 1T parameters. This section details the architectural specifics of this platform.

CPU Architecture#

DGX Station features one NVIDIA Grace CPU with 72 ARM cores based on the Neoverse V2 architecure. The CPU cores support the ARM v9.0 instruction set with crypto extensions and contain four 128b SVE2 units. The CPU has a 6-way instruction decoder, is able to dispatch 8 instructions per clock and has 6 scalar ALUs. Each core includes 64 KB of L1 instruction cache and 64 KB of L1 data cache, as well as 1 MB of L2 cache (8-way set associative). The CPU provides a total of 114 MB of L3 cache and connects to up to 496 GB of LPDDR5X memory.

For more information, refer to ARM CPU Architecture, ARM Memory Ordering and the Grace CPU architecture description.

GPU#

The GB300 Desktop Superchip contains one Blackwell Ultra GPU with fifth-generation Tensor Cores and support for NVFP4 computation. The GPU has up to 252 GB of HBM3e memory, delivering up to 7.1 TB/s of GPU memory bandwidth. It supplies the system’s primary acceleration for AI and other CUDA workloads. The CPU and GPU are connected with a high speed NVLink C2C interconnect that facilitates coherent access to CPU and GPU memory.

System Architecture: Memory and Interconnects#

DGX Station combines two physically distinct memory resources in the GB300 coherent memory domain. These are LPDDR5X attached to the Grace CPU and HBM3e attached to the Blackwell Ultra GPU. NVLink-C2C connects the processors and supports coherent access between them.

Memory and Interconnect Specifications#

Component

Capacity

Peak Bandwidth

CPU memory: LPDDR5X

496 GB

396 GB/s

GB300 GPU memory: HBM3e

252 GB

7.1 TB/s

CPU-GPU interconnect: NVLink-C2C

Not applicable

900 GB/s (bidirectional)

The combined CPU and GB300 GPU memory capacity is up to 748 GB. The NVLink-C2C figure describes aggregate bidirectional link bandwidth.

Memory locality affects application performance. GPU access to CPU-attached memory traverses NVLink-C2C and has different performance characteristics from access to local HBM3e. Coherent access therefore does not mean that the two memory resources are interchangeable in terms of performance.

Networking#

DGX Station includes an NVIDIA ConnectX-8 SuperNIC with two 400 Gb/s QSFP112 ports, providing up to 800 Gb/s of aggregate Ethernet connectivity.

Software#

DGX Station is preinstalled with Ubuntu with NVIDIA AI Developer Tools which is based on Ubuntu version 24.04. NVIDIA drivers, libraries, frameworks and tools are included.