One of Amazon's most recent announcements was the release of their new tool called Amazon Rekognition Custom Labels. This advanced tool has the capability to improve machine learning on a whole new scale, allowing for better data analysis and object recognition. Amazon Rekognition will help users train their machine learning models more easily and allow them to understand a set of objects out of limited data. In other words, this capability will make machines more intelligent and capable of recognizing items with far less data sets than ever before. Employees stand near an The Amazon Inc. logo is displayed above the reception counter at the ... [ ] company's campus in Hyderabad, India, on Friday, Sept. 6, 2019.

I'm on mobile, but I checked the source and kernels provided. I don't understand why they have no GPU abstraction layer or at least a bettter tuned CUDA derivative. Their project seems to only support NVIDIA graphics cards, but with the increasing diversity in the GPU market, especially enterprise offers, it would've been a good sub-project. Maybe they cut energy costs, reduce node utilization by using special purpose GPU's and load-balancing. Similar to the BigLittle architecture found in ARM CPUs.