Talking about machine-vision (AI algorithms), the best comes with
NVIDIA Tegra K1 SoC and uses the same Kepler computing core designed into supercomputers around the world. This gives you a fully functional
CUDA platform for quickly developing and deploying compute-intensive systems for computer vision, robotics, medicine, and more.
Do you expect to find them in hospitals, quality control production, ... ? You will have a surprise.
I have recently seen an old SGI system(1) made in the 90s for quality control in production. Things like you assemble to components on the PCB with a pick & place machine, then it's wave soldered. Before handing the finished product to the customer, the assembly company will check the finished product. The can be done with X-ray but also using a camera and OCR to make sure components are placed correctly. Some things cannot be automated and have to be done manually. It also depends on the value of the card and the size of the production run.
Shocked? Read the following
Yes, and the most extreme example of this I dealt with was a textile factory in Turkey that bought some systems from me way back. Their main knitting machine, a huge one for constructing very wide carpets, would have had to be replaced if they'd upgraded to the newer PC-based platform for managing the textile pattern files, costing a quarter of a million Euro. It was enormously cheaper just to upgrade their existing O2s and buy some more powerful maxed-out Octane2s.
A hospital in France told me something similar, replacing their SGI system with a PC for controlling a medical scanner would have required legal recertification of the scanner, taking many weeks and costing 50K Euro. Simply replacing the Indigo2 with another unit was a 200x cheaper.
Things do eventually get retired of course, but it can take a long time. A UK aerospace company I know has only just stopped using its R3000 IRIS Indigo systems, whereas a US company is still using a 12MHz Personal IRIS they bought from me for industrial process control. From a hospital in Barbados to a factory in Hiroshima, or even a small woollen clothing shop in Edinburgh (billbaber.com), SGIs are still used all over the place, lurking behind the scenes
(1) MIPS { R4K, R5K, R10K, R12K, R14K, R16L } CPU + OpenGL v1-v2 in hardware via dedicated GFX, large use of the NUMAflex and crossbar matrices to connect the CPUs to devices.