One Ware & Altera white paper for defect detection with MAX 10 FPGA

In a just released white paper (link), One Ware explains how they implemented a tailored AI model (CNN) using only a fraction of the logic capacity of the chip. They compare the performance with a raw approach using Jetson nano GPU from Nvidia and claim major interest in power consumption, accuracy and cost.

Although specific conditioning of the AI model is mandatory (Quantized Aware Training) and clever dimension reduction is applied at different stages, it is still remarkable to obtain such results with super limited ressources. See the following Table.

Last but not least, the authors explain how the widespread use of MAX10 devices in factories is due to their assets for industrial machinery. We understand that the ability to deploy more performant defect detection algorithms on the same hardware is a very strong driver. Indeed it minimizes changes in factory organization (space, power consumption, data management) and is accessible at a very low cost.

Beautiful work!