# Intelligent Delivery: Optimizing Compute for Large-Scale Ads Ranking Meta implements an ML-driven system to dynamically allocate compute resources for ads ranking. - **Source:** [Engineering at Meta](https://brifing.si/source/28383801-e365-4f15-8f20-49d8c0635946) - **Published:** 2026-11-11 - **Topic:** [technology](https://brifing.si/c/technology) - **Original article:** - **On Brifing:** https://brifing.si/s/a09ca3a9-b1f8-412d-80c6-decb46973812 ## Scaling Ads Ranking Systems Ads ranking models are rapidly expanding toward LLM scale, creating significant pressure on existing hardware infrastructure. Because compute resources do not scale at the same rate as model complexity, Meta developed Intelligent Delivery. This system optimizes resource allocation across the serving stack to ensure that every unit of compute serves ads effectively. ## System-Wide Performance Gains Intelligent Delivery has achieved approximately 10% incremental ranking performance improvements since 2017. By utilizing an ML-driven adaptive layer, the system continuously directs resources toward requests, models, and hardware fleets with the highest marginal utility. This approach ensures that advanced ranking models operate efficiently within strict latency budgets and finite serving capacity constraints. ## The Challenge of Heterogeneity The marginal impact of compute is highly context-dependent, varying significantly based on product surface, user intent, and system state. Simple rule-based segmentation fails to capture nonlinear interactions among thousands of signals. Consequently, Meta shifted toward ML-driven predictors to estimate the return on every tunable system parameter for each individual ad request. ## Resource Core Model Architecture The Resource Core Model serves as a shared intelligence layer, featuring 6B sparse and 20M dense parameters. It runs before candidate retrieval to capture nuanced request patterns through learned embeddings. By fine-tuning this model, engineers generate incremental signals regarding outcome, compute consumption, and latency impact, enabling more precise resource allocation decisions across the funnel. ## Hierarchical Decision Engines Meta decomposed the resource distribution problem into a hierarchy of decision engines to manage its globally distributed, heterogeneous fleet. This structure addresses three critical questions: which region serves the request, which hardware it lands on, and how much compute the host allocates. These layers ensure that scarce hardware is assigned to the most impactful workloads. ## Regional Traffic Optimization Traffic and capacity are not colocated, and each data center region maintains a different hardware mix and headroom. Meta built real-time infrastructure to shift traffic between regions in cycles of minutes. This mechanism drives the system toward a state where every region achieves equal efficiency on its marginal unit of capacity, balancing the global load. ## Hardware-Specific Workload Matching The fleet includes diverse CPUs and accelerators, such as NVIDIA, AMD, and Meta’s own MTIA chips. Because capacity is not interchangeable, Meta benchmarks efficiency for every workload-SKU pair. This allows the system to solve assignment problems, ensuring that specific hardware generations are matched with the workloads that convert them into the most useful work. ## Host-Level Latency Management Within each host, concurrent requests compete for resources under a hard tail-latency SLO. The system predicts expected outcomes and costs in real time, solving constrained optimizations to maximize impact. As host load increases, the framework trims resource-intensive, low-impact work, enforcing decisions from the kernel-level scheduler up to application-level thread pools to maintain performance. ## Agentic AI Optimization Loop Manual configuration of millions of tunable dimensions is unsustainable for human engineers. Meta now employs AI agents that monitor the system landscape, reading source code and experiment logs to propose changes. In the first half of 2026, these agents contributed to 20% of optimizations, shortening the development loop from months to mere weeks. ## Automation and Human Oversight On the most automated mechanisms, AI agents now handle 80% of the engineering effort, including hypothesis formation and experiment execution. Despite this high level of automation, human engineers retain control by reviewing and approving every proposed change. This partnership captures institutional expertise while allowing the system to adapt rapidly to shifting optimization landscapes. ## Future Technical Frontiers Meta is currently working toward end-to-end real-time optimization by fusing request understanding and serving into a single model. Additionally, the team is reimagining levers for GPU-based serving, as large models like the Generative Ads Model require different throughput and memory management strategies compared to traditional CPU-based ranking architectures used in previous years. ## Sources covering this story - https://engineering.fb.com/2026/11/11/ml-applications/intelligent-delivery-serve-better-ads/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._