Machine vision computer: how to choose the GPU, cameras and form factor

What the computer does in a vision system

It receives the images from the cameras, prepares them (cropping, light correction, alignment), runs the algorithm or the AI model and returns a decision: good or defective part, a measurement in or out of tolerance, a code read or not. That decision goes to the line’s PLC, which rejects the part or stops the machine.

All of this happens within a maximum time per part, set by the line speed. That is why the computer is not chosen by the longest spec sheet, but by what it has to do within that time, with every camera at once and over whole shifts.

The GPU: size it by the model, not by the catalogue

If the system uses deep learning (classifying, detecting or segmenting defects), inference runs on the GPU. What limits it is its memory (VRAM): it has to hold the model, the images processed at once and, if there are several models, all of them.

As a guide: a detection or classification model with one or two cameras runs well on an 8 to 16 GB card; several high-resolution cameras, segmentation or several models at once call for 16 to 32 GB. On lines that run around the clock, RTX PRO cards bring ECC memory, moderate power draw (the RTX PRO 2000 Blackwell uses 70 W and is low profile) and single-slot versions for compact enclosures.

Cameras and interfaces: each one needs a different port

GigE Vision is the most common in industry: cameras over network cable, often powered by PoE. A network card with one port per camera (or a dedicated switch) and jumbo frames enabled keep images from competing with other traffic.

USB3 Vision is simple and cheap, but every camera needs its own bandwidth: independent USB controllers are better than a hub. CoaXPress and Camera Link use a frame grabber in a PCIe slot, so you need to count free slots and PCIe lanes on top of the GPU’s.

Processor, memory and disk

The processor prepares the images, manages the cameras and runs classic vision (measurements, templates, code reading), which often needs no GPU. With several cameras, more cores help. Memory: 32 to 64 GB depending on the number of cameras and the image size.

Disk depends on how much you keep: a 5-megapixel mono camera at 20 frames per second produces about 100 MB/s. Keeping everything calls for fast NVMe and plenty of capacity; the usual approach is to keep only rejected parts and a sample of good ones, to retrain the model.

Form factor: cabinet, machine or desk

Without deep learning, a fanless industrial PC goes inside the electrical cabinet, on a DIN rail or on the machine itself, with no moving parts to fill with dust. With a GPU, a compact unit with filtered airflow or a tower in the control cabinet. If the environment is demanding (vibration, dust, humidity), an industrial range tested to MIL-STD-810H.

Also check the power supply (many cabinets provide 24 V DC), automatic start after a power cut, and that the machine is tested inside its final enclosure, not on a bench.

Link to the PLC and the plant

The decision has to reach the PLC: through digital I/O (to trigger the camera and reject the part) or an industrial bus such as Profinet, EtherNet/IP, Modbus TCP or OPC UA, depending on what the line uses. If needed, an I/O card or a second dedicated network port is added.

On top of that, a watchdog that restarts the machine if it hangs and a frozen system image, so every machine in the plant is identical and can be swapped in minutes.

Indicative starting points

If the system is your product: the same machine for years

When the vision system is what you sell, the problem changes: you need the computer to be the same in unit one and unit one hundred, for years, with your image loaded and a traceable serial number. That is what the OEM programme is for: a frozen platform, a 24-hour test per unit and immediate replacement.