Our line needs 4 GigE cameras at 5 MP for visual inspection, using existing imaging software. What CPU should we choose, and is a discrete GPU mandatory?

Three things decide the configuration: camera bandwidth, algorithm type, and line cycle time. Bandwidth: four 5 MP mono cameras at about 20 fps produce close to 400 MB/s. One Gigabit port carries roughly 110 MB/s, so you need a multi-port NIC (or a 4-port PoE card) rather than one shared port, and the camera network should be separated from the plant network. Algorithm: for measurement, code reading or defect detection with classical vision, a current 4–8 core CPU (Core i5/i7) is enough and no discrete GPU is needed. A GPU only becomes necessary for deep-learning surface classification — then choose a machine with a PCIe x16 slot and adequate power, or one with an integrated Jetson module. Cycle time: you need to know how many milliseconds each part allows for processing. That number decides whether you must run several threads in parallel, which affects core count more than clock speed. Also budget spare RAM (5 MP frames times the number of threads), use an industrial SSD, and keep one expansion slot free for future I/O cards.