How Meta Checks More Ads Without the Same Growth in Costs

Meta uses Andromeda to find ads that may interest each person. This system first chooses possible ads from a very large group. Another system then looks at that smaller group in more detail. The first choice matters because good ads must reach the next step. Meta wants to improve this choice without making costs grow as quickly.

There are now many more ads for the system to check. Tools that make pictures and other material have helped create them. Meta changed its model to deal with this growing number of ads. A model is a computer tool that learns from information. Different parts now study different interests and changes in people's behaviour. The system chooses which parts can best help judge each ad.

Meta also changed how information moves through the computers doing this work. Sending large amounts between different machines had become a major limit. The new design puts more steps together on one GPU. A GPU is a computer chip that handles many calculations together. This cuts down the information that must travel between machines. Meta reports that the full system checks over one billion ads per second.

That figure describes the full process running on a single GPU server. It is different from faster results for just one individual step. Meta worked with NVIDIA to test and improve those separate steps too. The company says there is still room for further growth. Its main lesson is to design models, software, and hardware together. Better ad choices should not depend only on adding more computers.