Intelligent Delivery: Making Every Unit of Compute Serve Better Ads
AI Summary
Meta's Intelligent Delivery system uses machine learning and AI agents to dynamically allocate finite compute resources across a complex, heterogeneous ads ranking stack. By predicting the marginal utility of compute for individual ad requests and adjusting resource distribution across regions, hardware fleets, and hosts, the system achieves a global optimum and delivers significant ranking performance gains.
Key Takeaways
- The Resource Core Model utilizes deep learning with 6B sparse and 20M dense parameters to predict request-level marginal utility before candidate retrieval.
- Multi-layered mechanisms handle dynamic resource allocation across regional traffic routing, heterogeneous hardware SKUs, and host-level thread pools under strict tail-latency SLOs.
- AI agents assist in continuous system optimization by analyzing logs, proposing configuration changes, and designing experiments, contributing to 20% of optimizations in early 2026.
Why it matters
As ads ranking models grow to LLM scale, hardware capacity cannot keep pace with model complexity. Intelligent Delivery provides a blueprint for operating massive ML workloads efficiently under tight latency budgets and heterogeneous hardware constraints, maximizing system throughput and performance without requiring a linear expansion in physical compute infrastructure.
Conversation 0
Be the first to share your perspective.