⚡Meta Andromeda Now Scores Over One Billion Ads Per Second

Your ad targeting just got a lot faster

Meta just pushed its ads retrieval engine into territory most teams never plan for: over 10 billion ads scored per second on a single GH200 GPU. Andromeda's newest iteration scores every eligible ad with near ranking-level model quality, and it surfaces top candidates at over one billion ads per second. That is a new throughput benchmark for retrieval-stage models. Why should you care? Because the real lesson is about where the wall is. The 2024 version ran at tens of millions of ads per second per card. At billions per second, moving ad tensors through the memory hierarchy becomes the bottleneck, not the math. If you're tuning inference for recommendation or ranking systems, that's the part to study. Meta's answer is co-design: one GPU pipeline replacing fragmented, distributed workloads. The architecture borrows mixture-of-experts from LLMs and uses it for retrieval. Meta says this sets a new performance-efficiency Pareto frontier, with model capacity growing without a proportional rise in GPU cost. The 2026 iteration fuses model and system work into a single kernel per user-ad interaction, co-locating per-user request logic on one GPU box. Meta hasn't published cost figures, so the efficiency claims are still vendor numbers.