Andromeda: A New Era of Speed and Scale in Meta Ads Retrieval

Feb 06, 2026

Andromeda: A New Era of Speed and Scale in Meta Ads Retrieval

Introduction

Andromeda goes even beyond what Advantage+ automation can do by enhancing it. Combining advanced machine learning techniques with large-scale, real-time systems, Andromeda is Meta’s most advanced AI-driven ad retrieval engine, designed to operate at unprecedented scale and personalization. It is already used to define, personalize, and serve ads all around Facebook and Instagram to enormous audiences.
Because retrieval defines which ads are even eligible to be considered, Meta developed Andromeda, a single, AI-driven retrieval engine capable of scaling far beyond traditional back-end systems.

Why Relevance For Retrieval Has Never Been So Important

Meta carries out the delivery of ads in several steps, but retrieval is the starting point. Meta automates the process of sifting through millions of ads, selecting a relevant subset for subsequent ranking and display. Every tap, every play, and every scroll adds signals that increase the size and diversity of the advertising inventory.
Meta has seen Advantage+ automation significantly increase the size and diversity of its ad inventory. This growth is driven by automation across budgets, audiences, placements, and creatives, while generative AI further accelerates retrieval and decision-making. At this scale, using traditional approaches is no longer feasible because these methods slow down processes, consume immense resources, and limit the creation of tailored advertising experiences.

Consumer expectations for relevance and speed continue to rise.

Advertisers increasingly expect higher ROAS without spending excessive time manually restructuring campaigns, optimizing delivery, or managing real-time bidding for custom audiences. Even minor delays or irrelevant ad experiences can alienate users over time and negatively impact overall performance metrics. Andromeda was designed to directly address this challenge.
Andromeda is successful because of the seamless integration of algorithms, infrastructure, and system design.

What makes Andromeda different

Andromeda utilizes deep neural networks optimized for the latest hardware advancements. Its models are designed from the ground up to take full advantage of next-generation hardware platforms, including NVIDIA Grace Hopper–based systems.
Rather than adapting hardware-agnostic models to new infrastructure, Andromeda follows a co-design approach where models and hardware are built to work together.

Hierarchical indexing at scale.

When managing millions of ad creatives, merely having available hardware isn't sufficient; hierarchically indexing ads is. Hierarchically nesting ads into larger groups enables the system to shift the majority of computational bid processing from individual items to the super-sets of items. This hierarchical structure enables sub-linear inference costs, allowing the system to scale without a proportional increase in compute resources. This type of system maintains its performance as the number of ads grows without an outright increase in system resources.

Utilize model elasticity to enhance efficiency and return on investment (RoI).

The ability to apply the right level of model complexity, depending on the context, aka model elasticity, is core to Andromeda. The architecture is segment-aware, allowing it to increase or decrease computation more effectively in an intelligent way. These savings enhance efficiency and allocate funds more effectively, helping Meta maintain a positive return on investment in infrastructure while ensuring advertiser and customer satisfaction.

Optimized Retrieval Process

The ingest loop and retrieval process have been optimized to reduce computational and memory overhead. GPU-aware processing and optimized software tools and libraries can deliver feature extraction speeds that are up to 100 times faster compared to CPU-based high-throughput feature extractors and 3x end-to-end inference throughput. This advancement makes true real-time personalization feasible at a massive scale.

Conclusion

The Andromeda isn’t merely a performance upgrade; it’s a complete conceptual overhaul of what a structured advertising system for the masses should be. With an AI-focused design, up-to-date hardware, and wise use of retrieval, Meta is creating a new standard for personalized advertising. A paradigm where system intelligence amplifies strong creative, rewarding originality with greater visibility and performance.

Q

Tell us how can we assist you?

We are always happy to answer any questions!