Meta's Intelligent Delivery System Delivers 10% Incremental Ads Ranking Performance via ML-Driven Compute Allocation

Details

Meta's Intelligent Delivery system allocates compute resources across the serving stack to achieve a system-wide optimum, contributing ~10% incremental ads ranking performance on top of model and system advancements from 2017 through the first half of 2026.
The system addresses the challenge of scaling ads ranking models toward LLM scale by using ML to predict the marginal utility of compute, as extra compute has varying impact depending on the request, model, and hardware.
A Resource Core Model with 6B sparse parameters and 20M dense parameters identifies high-impact ad requests, achieving a ~20% improvement in identifying these requests in offline benchmarks compared to previous production models.
Agentic AI now contributes to about 20% of Intelligent Delivery optimizations, with agents handling up to 80% of engineering effort on the most automated mechanisms, shortening the optimization loop from months to weeks.
Allocation mechanisms operate at three levels: shifting traffic between regions to equalize marginal efficiency, assigning workloads to specific hardware (Intel, AMD, Arm, NVIDIA, MTIA) based on measured efficiency, and optimizing per-host resource usage under latency SLOs.
Future work focuses on fusing request understanding and serving into a single real-time model and adapting optimization levers for GPU-based serving of large models like Meta's Generative Ads Model (GEM).

Key point