Skip to content
EMARQUE AI

Workstation, server or rack-scale?

Four sizes of on-prem AI system, with different power, cooling and installation requirements. The room usually constrains the choice more than the GPU does, so this guide starts there.
Published . Updated .
Size comparison

The four sizes, side by side.

Personal, desk-side

One developer or researcher, or a small team, with the NVIDIA stack on the desk. NVIDIA DGX Spark and the EMARQUE AI Work 100.

Model size
DGX Spark: up to 200B parameters (NVIDIA figure). AI Work 100: 70B class at 4-bit on one 48 or 96 GB card, at 16-bit across two 96 GB cards.
Form
Compact desktop (DGX Spark) or desk-side tower (AI Work 100)
Cooling
Office air, no special HVAC
Power
DGX Spark 240 W; the AI Work 100 runs from a standard wall outlet
Step up
The model no longer fits the memory you have, or a department needs it rather than one desk.

Workstation

NVIDIA DGX Station, a Grace Blackwell Ultra machine that installs beside a desk, and the EMARQUE AI PRO 500, four RTX PRO 6000 GPUs on a Threadripper PRO workstation platform in a 5U chassis.

Model size
DGX Station: 748 GB of coherent memory (252 GB HBM3e + 496 GB LPDDR5X), so large models run on one machine. AI PRO 500: 384 GB of GPU memory across four cards
Form
Desk-side tower (DGX Station) or 5U rackmount (AI PRO 500)
Cooling
DGX Station: cooled inside the unit, in a normal office. AI PRO 500: air-cooled, in a rack
Power
DGX Station 1,600 W, on a 20 A circuit (NVIDIA figure); AI PRO 500 on a 3,000 W supply
Step up
You need more than one machine, or it has to live in a rack with the rest of IT.

Rackmount server

Single-node rackmount platforms: the EMARQUE AI Server (six or eight GPUs) and AI HGX B300, plus NVIDIA DGX B200 and DGX B300.

Model size
70B-class serving, fine-tuning, or several models side by side
Form
6.5U to 10U rackmount
Cooling
Air-cooled, in a server room or with in-row cooling. The AI HGX Rubin is liquid-cooled
Power
Up to about 14 to 14.5 kW for an eight-GPU HGX or DGX system (NVIDIA figures for DGX B300); 9,600 W of supplies in the eight-GPU AI Server
Step up
One node is no longer enough and the work has to spread across a rack of GPUs on one NVLink fabric.

Rack-scale

NVIDIA GB300 NVL72 and Vera Rubin NVL72, where a whole rack behaves as one accelerator. We source these through NVIDIA's OEM partners and coordinate the install.

Model size
Training and serving the largest models; 20 TB of HBM3e per GB300 NVL72 rack
Form
One rack as a single accelerator (NVL72)
Cooling
Direct-to-chip liquid, with a coolant distribution unit
Power
Confirmed with the OEM at design. NVIDIA publishes no per-rack figure.
Step up
More racks on the same fabric.

Tell us about your workload.

Tell us what the system has to do, how many people will use it and where it will run. We recommend a size before anything is specified.