Meta has announced Andromeda, its proprietary machine learning system for personalised ad retrieval.
The system establishes an efficient scaling law for retrieval with the aid of AI, driving a step-function improvement in value to advertisers and people.
Andromeda features a deep neural network custom designed for the NVIDIA Grace Hopper Superchip to deliver superior retrieval performance.
It utilises a hierarchical index to scale the volume of ad creatives, empowering the adoption of GenAI technologies by advertisers.
Andromeda reduces system complexity by minimising components and rule-based logic, allowing for end-to-end performance optimisation.
The system's design is optimised for AI hardware, minimising memory bandwidth bottlenecks and enabling highly parallel, computation-intensive models with high performance.
Andromeda represents a significant technological leap in retrieval – addressing volume and latency challenges through key ML and system innovations.
Meta's investment in AI aims to deliver the best experience given the strict latency and capacity ROI budgets and drive better advertiser outcomes.
Integrating Andromeda with MTIA and future generations of commercially-available GPUs will further improve advertiser performance and push the boundaries of scaling retrieval.
Andromeda has been designed to maximise ads performance and has achieved a 6% recall improvement to the retrieval system, delivering an 8% ads quality improvement on selected segments.