The Pulse
Meta Rebuilds Andromeda to Score a Billion Ads per Second
Meta says its rebuilt Andromeda ads retrieval system can score more than one billion eligible ads per second on a single GPU server. The redesign combines a mixture-of-experts model with a fused GPU pipeline, and Meta attributes a 2.4% Face
AI.info Team · Written with AI · fact-checked ·
Meta says its rebuilt Andromeda ads retrieval pipeline can score more than one billion eligible ads per second on a single GPU server, a throughput figure the company describes as a new benchmark for retrieval-stage models. In a post published October 28, Meta’s engineers say the system pairs a new mixture-of-experts model with a GPU pipeline that handles ad matching, scoring and candidate selection in one place.
Andromeda operates at the first stage of Meta’s ad recommendation process. It narrows millions of eligible ads to a smaller pool in milliseconds; ranking models then examine those candidates in more detail. Meta says the new system brings retrieval model performance close to ranking-level quality while keeping the same latency budget.
Andromeda’s MoE model cuts inference work
The new model replaces a single user-modeling tower with an ensemble of specialized user experts. It draws on signals including ad engagement sequences, user-cohort representations, semantic features produced by large language models and interactions across Meta services. A gating network routes each candidate ad to the most relevant experts, while the ad representation also incorporates real-time serving context.
Meta says the model approaches ranking-level performance with about 300 times fewer inference-time floating-point operations than its 2024 Andromeda model. On a single NVIDIA GH200 GPU, the model’s inference throughput exceeds 10 billion ads per second. That is a model-level throughput figure; the production system’s reported rate is lower because it also performs other retrieval tasks.
One GPU now handles matching, scoring and selection
In the older design, GPU model execution sat alongside other work spread across CPU hosts and remote accelerators. That arrangement moved large candidate lists across network links and generated substantial memory traffic between GPU operations. At the new target throughput, Meta says moving data—not the arithmetic—became a key constraint.
The redesigned pipeline fuses model operations and system work into a single GPU kernel. It also moves index matching, scoring and top-K selection—the process of keeping the best candidates—onto the same GPU server. Rather than sending large lists between machines, the system sends a compact request containing user features and a targeting query, then returns the selected ads. Meta says this reduces network traffic by about 100 times.
“The 2026 Meta Andromeda pushes the model-system co-design to the next level with two innovations — a single GPU kernel fusing both model and system workloads per user-ad interaction and a strategy to co-locate per user request logics with per user-ad GPU kernels in a single GPU box.”
Haiyu Lu, Shuo Gu, Yang Yang (Tony), Jingwen Ren, Tianchen Ji, Qing Zhang, Wentian Guo, Haocheng Wu, Siqi Zeng, Yunchen Pu, Liang Wang and Chonglin Sun, Meta
Meta and NVIDIA report faster kernels
Meta says it worked with NVIDIA on GPU optimizations, with each component in the collaboration achieving between twofold and fivefold speedups over its baseline. The optimized mixture-of-experts gating kernel rose from 2.4 billion to more than 10 billion ads per second, reaching about 85% of the GPU’s peak high-bandwidth memory performance. Meta also reports that INT8 quantization delivered up to twice the throughput of its optimized FP16 kernels, with accuracy on par with 16-bit precision.
Index matching reached more than 100 billion ads per second in the reported tests—over ten times the scoring network’s rate. The full pipeline, after the additional optimizations, benchmarks at multiple times the current production throughput on one GPU server. Meta says it still expects to integrate further gains into production, so the benchmark should not be confused with a claim that the full end-to-end system already runs at that higher rate.
The conversion figure is Meta’s own claim
Meta says Andromeda has delivered a 2.4% increase in conversions on Facebook since 2025. The company presents that figure alongside the system’s ability to search a larger candidate pool within the same latency budget, but the engineering post does not provide a measurement method or comparison details for the conversion result.
The engineering team says it is extending its work to NVIDIA’s Grace Blackwell platform, including tests of how CPU-to-GPU resource ratios affect performance. For now, the concrete production claim is that Andromeda scores every eligible ad in its candidate pool at more than one billion ads per second on a single GPU server.