Meta Andromeda is Meta’s ad-retrieval system: it narrows eligible ads to candidates before later models rank them in more detail. In an engineering post published October 28, 2026, Meta said its newer, co-located production pipeline scores more than 1 billion eligible ads per second on one GPU server. The company also reported faster rates for individual components, which measure different parts of the system.
In this article
- What Andromeda does before an ad is ranked
- A retrieval model and its GPU pipeline
- Why the three throughput figures differ
- Meta’s reported results measure different outcomes
What Andromeda does before an ad is ranked
Ad retrieval is the first selection step. It filters eligible ads into a smaller candidate set; later ranking models then evaluate those candidates in greater detail. Andromeda is that retrieval stage, not the final ranking system. Meta’s Advantage+ suite is also distinct: it is an advertising automation suite, not another name for Andromeda.
A retrieval model and its GPU pipeline
Meta’s newer retrieval model uses a Mixture-of-Experts (MoE) architecture. Specialized user towers represent different user signals, while ad representations include real-time serving context. An attention-based gating network routes candidates to relevant user experts. Meta says this model approaches ranking-level performance with about 300 times fewer inference-time FLOPs than its 2024 retrieval model. FLOPs measure the arithmetic operations a model performs.
The system combines model calculations with the work around them. A fused GPU kernel gathers embeddings, calculates user-ad interactions and scores candidates. Intermediate results stay in GPU registers and shared memory rather than being written to high-bandwidth memory. Meta also places index matching, scoring and top-K selection on one GPU server. Index matching applies advertiser targeting rules; top-K selection returns the highest-scoring candidates. Meta says the co-located design reduces network traffic by about 100 times.
Meta says its joint optimization work with NVIDIA began in early 2026. NVIDIA contributed optimized GPU kernels, while Meta integrated and benchmarked the pipeline.
Why the three throughput figures differ
Each rate covers a different system boundary. The full production-pipeline figure includes more than the retrieval model alone; the other rates describe a model on one GPU or a single component.
| System boundary | Throughput reported by Meta | What it covers |
| Co-located production pipeline | More than 1 billion ads per second | Full pipeline on one GPU server, scoring eligible ads |
| Retrieval model | More than 10 billion ads per second | Model on one NVIDIA GH200 GPU |
| Index matching | More than 100 billion ads per second | Optimized index-matching component |
These figures describe separate parts of Andromeda, so the component rates do not represent the complete production pipeline. Meta also reported that an optimized gating kernel ran 5.3 times faster, rising from 2.4 billion to more than 10 billion ads per second, and used about 85% of peak high-bandwidth-memory bandwidth. For another optimization, Meta said INT8 quantization delivered up to twice the throughput of optimized FP16 kernels, with accuracy on par with 16-bit precision.
Meta’s reported results measure different outcomes
Meta’s 2024 account of Andromeda’s deployment across Facebook and Instagram reported a 6% increase in retrieval recall and an 8% improvement in ad quality on selected segments. Those are separate measures of retrieval and ad quality. In its October 2026 post, Meta attributed a 2.4% increase in Facebook conversions since 2025 to Andromeda. That figure concerns conversions over a different period, not the 2024 recall or selected-segment quality measures.
