Analysis
Meta’s October 28 engineering report describes a new generation of Andromeda, the retrieval system that selects candidates for later advertising rankings. The company reports production throughput above one billion ads per second on a single GPU server. That stage matters because an ad excluded from the candidate set cannot benefit from a more expensive ranking model later. Retrieval quality therefore sets a consequential boundary for the rest of the pipeline.
The redesign couples a mixture-of-experts model with a pipeline that brings index matching, scoring and final candidate selection together on the GPU. Meta says keeping intermediate work close to the computation reduces memory and network traffic. Its collaboration with NVIDIA includes faster kernels and system tuning. This is a coordinated engineering account, rather than evidence that changing the model alone produced the reported improvement.
The distinction between a component and a complete service is essential to reading the numbers. Meta reports more than ten billion ads per second for the fused scoring kernel, alongside the lower production pipeline figure. A kernel benchmark leaves out work that an actual request still needs. Operators comparing designs should ask which stages, transfers and latency constraints a number includes before using it to estimate required hardware.
There is a broader design lesson here: inspect where information travels before assuming the next bottleneck calls for more computation. That is an engineering inference from the architecture, not a promise of transferable savings. We have not reproduced Meta’s measurements or evaluated ad relevance. Its account supplies a detailed design to study, while the economics and user outcomes of another deployment would need their own evidence.
Source: Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design · 2026-10-28
Why it matters
End-to-end service performance depends on data movement and all pipeline stages, not only the fastest kernel.