GPU catalogue · NVIDIA
NVIDIA's deskside AI computer: one Grace Blackwell Ultra superchip with 748 GB of coherent memory, 252 GB of it fast HBM3e on the GPU. NVIDIA designs it, and Dell, HP, ASUS, MSI, GIGABYTE and Supermicro build and sell it.
Specifications
Good for: Running models up to about a trillion parameters at 4-bit on one machine, serving a team from one box, and fine-tuning models too big for any single card.
| Machine | NVIDIA DGX Station (GB300) |
|---|---|
| Memory | 252 GB HBM3e on the GPU + 496 GB LPDDR5X on the CPU (748 GB coherent) |
| GPU | 1× Blackwell Ultra, 7.1 TB/s memory bandwidth |
| CPU | Grace, 72 Arm Neoverse V2 cores, 396 GB/s memory bandwidth |
| CPU to GPU | NVLink-C2C, 900 GB/s |
| AI compute | Up to 20 PFLOPS FP4 (15 without sparsity) |
| Networking | ConnectX-8 SuperNIC, 2× 400 Gb/s |
| GPU partitions | Up to 7 (MIG) |
| Power | 1,600 W, needs a 20 A circuit |
| Form | Liquid-cooled deskside tower |
| Software | Ubuntu 24.04 with NVIDIA's AI tools (DGX OS) |
| Architecture | Grace Blackwell Ultra |
| Released | 2026-03 |
| Price to buy | From about €89,000 (base builds, 2026-10) |
Community data
| Test | NVIDIA DGX Station (GB300) | Compared with | Source |
|---|---|---|---|
| gpt-oss-20b, 1 request | 530 to 536 t/s | RTX PRO 6000: 241 to 244 t/s · DGX Spark: 50 t/s | StorageReview, vLLM |
| Qwen3 Coder 30B A3B (FP8), 1 request | 310 to 319 t/s | RTX PRO 6000: 169 to 208 t/s · DGX Spark: 55 t/s | StorageReview, vLLM |
| gpt-oss-120b, 1 request | about 380 t/s | H200 NVL: about 200 t/s · RTX PRO 6000: about 170 t/s · DGX Spark: about 35 t/s | StorageReview, vLLM (read from the chart) |
| gpt-oss-120b, 128 requests at once | about 9,300 t/s in total | H200 NVL: 3,376 t/s · RTX PRO 6000: 3,064 to 3,384 t/s · DGX Spark: 500 t/s | StorageReview, vLLM |
| Llama 2 7B Q4_0, llama.cpp (our decode ladder) | about 300 t/s (estimate) | B200: 298 t/s · GH200: 312 t/s · RTX 5090: 290 t/s | llama.cpp scoreboard, no DGX Station entry yet |
StorageReview ran vLLM with 512 tokens in and 512 out, one run each, on the MSI and ASUS builds.
Who builds it
| Maker | Machine | Memory |
|---|---|---|
| ASUS | ASUS ExpertCenter Pro ET900N G3 | 252 GB |
| Dell | Dell Pro Max with GB300 | 252 GB |
| GIGABYTE | GIGABYTE W775-V10 | 252 GB |
| HP | HP ZGX Fury AI Station | 252 GB |
| MSI | MSI XpertStation WS300 | 252 GB |
| Supermicro | Supermicro Super AI Station (GB300) | 252 GB |
Models
language models · image models · video models · embedding models.