The model has to fit in GPU memory, with room left for the context of every person using it. At 4-bit, a model's weights come to roughly half a gigabyte per billion parameters; at FP16, about two. Settle this first: it decides the class of server more than anything else.
On a PCIe server such as our AI Server, each GPU works on its own, which suits several models side by side, each on its own GPUs. On an HGX or DGX system the eight GPUs are linked by NVLink, which suits one large model, or training, across all of them at once.
An eight-GPU Blackwell Ultra machine is a data-centre load, not an office one: three-phase power and matching cooling. A PCIe AI Server is air-cooled with front-to-back airflow. We review power, cooling and rack space with you before quoting.
Servers we build carry a three-year system warranty, a 90-day exchange and lifetime labour. HGX servers carry a three-year warranty, with failed parts replaced through the manufacturer's RMA. NVIDIA DGX systems carry NVIDIA's support contract. In every case we handle the claim here.
For the figures we ask for before sizing a system, see how to size an on-prem AI system.
| Server | GPUs | GPU memory | How the GPUs connect | Size and room |
|---|---|---|---|---|
| EMARQUE AI Server | 6 or 8 × NVIDIA RTX PRO 6000 Blackwell Server Edition | 576 or 768 GB (96 GB per GPU) | PCIe Gen5; each GPU works on its own | 6.5U, air-cooled; 9,600 W in the eight-GPU size |
| EMARQUE AI HGX B300 | 8 × NVIDIA Blackwell Ultra on an HGX B300 board | 2.1 TB HBM3e | NVLink, 1.8 TB/s per GPU (NVIDIA figure) | 8U, built to order; a data-centre load |
| EMARQUE AI HGX Rubin | 8 × NVIDIA Rubin on an HGX Rubin NVL8 board | 2 TB HBM4 (NVIDIA figure) | NVLink, 3.6 TB/s per GPU (NVIDIA figure) | Liquid-cooled, built to order; needs coolant distribution |
| NVIDIA DGX B200 | 8 × NVIDIA Blackwell | 1,440 GB HBM3e (NVIDIA figure) | NVLink, 14.4 TB/s across eight GPUs (NVIDIA figure) | 10U, air-cooled; about 14.3 kW (NVIDIA figure) |
| NVIDIA DGX B300 | 8 × NVIDIA Blackwell Ultra | 2.1 TB HBM3e (NVIDIA figure) | NVLink, 14.4 TB/s across eight GPUs (NVIDIA figure) | 10U; about 14 kW (NVIDIA figure) |
| NVIDIA GB300 NVL72 | 72 × NVIDIA Blackwell Ultra and 36 Grace CPUs | 20 TB HBM3e (NVIDIA figure) | NVLink across the whole rack | One liquid-cooled rack |
NVIDIA figures from NVIDIA's DGX, HGX and GB300 NVL72 pages. The AI Server, AI HGX B300 and AI HGX Rubin rows are our own specifications on those boards. Each product page has the full detail.
Tell us the models you want to run and the room it will sit in. We recommend a server in writing, in MYR.