What integrated graphics do well
Decoding and encoding video (several high-definition cameras or one 4K stream), driving three or four screens, running accelerated classic vision (measurements, code reading, tracking) and small artificial intelligence models: detection on one or two cameras at a few frames per second. With processors that have an NPU, part of the inference is offloaded to that dedicated block and power draw drops.
When a dedicated GPU is needed
When there is deep learning with several cameras at line speed, segmentation or several models at once; when the model needs more than 8 GB of its own memory; when there is 3D or real-time rendering; or when inference is continuous and latency rules. In those cases a low-power dedicated GPU multiplies capacity several times over.
| Case | GPU | Indicative power draw |
|---|---|---|
| Video, screens, classic vision, one small model | Integrated (and NPU if available) | 15 to 65 W for the whole unit |
| Detection on 2 to 4 cameras, medium models | RTX PRO 2000 Blackwell, 16 GB, low profile | 70 W |
| Several cameras with deep learning, 3D | RTX PRO 4000 Blackwell, 24 GB, single slot | 140 W |
| Large models that do not fit in a small card | 64 or 128 GB unified memory | 50 to 120 W for the whole unit |
Heat in 10 litres
A 70-watt GPU in a 10-litre case is cooled with a filtered fan; a 140-watt one, with designed airflow and the case tested closed. Above that, it is no longer a compact unit. If the unit must be fanless, a dedicated GPU is not an option: integrated graphics, NPU or unified memory.
GPU form factor: low profile and single slot
A compact unit takes low-profile (half-height) or single-slot cards, powered through the PCIe slot itself, with no extra cables. The RTX PRO 2000 and 4000 Blackwell exist in those formats and bring ECC memory and certified drivers; low-profile GeForce cards work for prototypes, but take two slots and lack the same supply continuity.
How to decide in five minutes
Count the cameras and the frames per second the process needs, measure the memory of the model you will use in production and ask whether the unit may have a fan. With those three answers you choose the GPU; the rest (processor, memory, ports) follows.
Related hardware and guides
Frequently asked questions
Does the NPU replace the GPU?
For small vision and speech models with low power draw, yes. For deep learning with several cameras or large models, no: its power and memory are far below those of a dedicated GPU.
How many cameras can integrated graphics handle?
It depends on the model and the frames per second. As a guide, one or two cameras with a detection model at a few frames per second. Beyond that, a dedicated GPU.
Can a dedicated GPU go in a fanless unit?
Not sensibly: a 70 or 140-watt GPU needs airflow. If the unit must be sealed, use integrated graphics, an NPU or unified memory, or move inference to a ventilated unit in the cabinet.
GeForce or RTX PRO in a compact unit?
RTX PRO: they exist in low-profile and single-slot formats, are powered through the slot, bring ECC memory and are manufactured for years. GeForce cards are fine for a prototype, not for a series.
Where will your computer work?
Tell us the location, the power available and what it has to do. We will propose the compact unit, with tests and a price, within 12 working hours.