Meta's Intelligent Ads Delivery: Making Every Unit of Compute Serve Better Ads (engineering.fb.com)

Meta has introduced an innovative system called Intelligent Delivery, designed to optimize ad delivery by efficiently allocating computational resources across its serving stack. As ads ranking models grow in complexity, the demand for computational resources has outpaced hardware advancements, leading to challenges in maximizing performance. Intelligent Delivery addresses this by leveraging machine learning to predict the marginal utility of compute for different ad requests, resulting in an approximate 10% improvement in ranking performance since its inception in 2017. This system incorporates advanced capabilities like a shared ML intelligence layer, which enhances understanding of ad requests, and a hierarchy of decision engines that manage resource allocation dynamically across varied hardware and data centers. Additionally, it employs agentic AI to continuously monitor and optimize the entire serving landscape, allowing for rapid adaptations to system changes, thereby shortening the optimization cycle from months to weeks. This approach not only improves operational efficiency but also ensures that ad delivery remains responsive to evolving user behaviors and technological advancements, promising significant implications for the future of AI-driven advertising strategies.