Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine

This post introduces Andromeda, Meta's next-generation personalized ads retrieval engine. It details innovations in ML model architecture, feature representation, learning algorithms, indexing, and inference paradigms, specifically designed to leverage NVIDIA Grace Hopper Superchip and MTIA hardware. Key technical contributions include: a deep neural network custom-designed for superior performance on the Grace Hopper Superchip, achieving +6% recall improvement and +8% ads quality improvement; hierarchical indexing to support exponential ad creatives growth from Advantage+ creative, enabling efficient scaling to a large volume of ads creatives; and AI development efficiency through system simplification. Andromeda represents a new paradigm in ads retrieval by employing a state-of-the-art deep neural network with sublinear inference cost, hierarchical indexing jointly trained with retrieval models, model elasticity for agile resource allocation, and an optimized retrieval model with enhanced parallelism and low-latency, high-throughput GPU operators. The system achieves over 100x improvement in feature extraction latency and throughput compared to previous CPU-based components and enhances end-to-end model inference QPS by over 3x. Future work includes transitioning to an autoregressive loss function and integrating with MTIA and future GPUs for further scaling.