--- title: Meta Andromeda: Powering the Next Generation of Ads... description: Meta's next-generation Andromeda ads retrieval system now scores every eligible ad and surfaces top candidates at over one billion ads per second, a 100x... canonical: https://daily.dev/posts/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-model-system-co-design-q0bvljfan twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design | daily.dev og:description: Meta's next-generation Andromeda ads retrieval system now scores every eligible ad and surfaces top candidates at over one billion ads per second, a 100x... og:url: https://daily.dev/posts/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-model-system-co-design-q0bvljfan og:image: https://api.daily.dev/og/posts/q0bVljfan.png og:image:alt: Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design **[Facebook Engineering ](https://daily.dev/sources/facebook_code)** · 12 min read · 0 upvotes · 0 comments ## Summary Meta's next-generation Andromeda ads retrieval system now scores every eligible ad and surfaces top candidates at over one billion ads per second, a 100x throughput jump over the 2024 system. A new mixture-of-experts retrieval architecture reaches near ranking-level model quality at about 300x fewer inference FLOPs. A fused GPU kernel and co-located index matching, scoring and top-K selection on a single GPU cut network traffic by roughly 100x. Joint work with NVIDIA pushed the gating kernel to over 10 billion ads per second per GH200 GPU, with INT8 quantization adding up to a further 2x. Since 2025, the system has driven a 2.4% increase in conversions on Facebook. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much throughput improvement did Meta get from its 2026 Andromeda ads retrieval system compared to 2024? The system achieved a 100x throughput lift over the 2024 version, now surfacing top ad candidates at over one billion ads per second on a single GPU server while scoring every eligible ad. The new mixture-of-experts gating architecture also reaches near ranking-level model quality using roughly 300x fewer inference-time FLOPs than the prior retrieval model. _Engineers scaling ML inference systems can follow model-system co-design patterns like this one on daily.dev._ ### How did Meta reduce network bottlenecks in its GPU-based ads retrieval pipeline? Meta co-located index matching, scoring, and top-K selection onto a single GPU server instead of shipping candidate lists across separate hosts. The system now sends only a compact request payload of user features and a targeting query, cutting network traffic by about 100x and shifting the bottleneck back to GPU compute rather than network bandwidth. _Developers designing distributed ML inference pipelines can track architecture trade-offs like these on daily.dev._ ### What throughput gains did Meta and NVIDIA achieve by optimizing GPU kernels for ads retrieval? The optimized interaction network (MoE gating) kernel improved 5.3x, from 2.4 billion to over 10 billion ads per second, reaching about 85% of peak HBM bandwidth. Index matching throughput improved over 2x to more than 100 billion ads per second. INT8 quantization added up to a further 2x gain over FP16 kernels with accuracy on par with 16-bit precision. _Teams evaluating GPU kernel fusion and quantization trade-offs can find deep dives like this on daily.dev._ ## Similar posts on daily.dev - [Meta’s Generative Ads Model \(GEM\): Scaling the Central Model and Compute Engine for Frontier Ads AI](https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd) · Facebook Engineering · 0 upvotes · 0 comments - [How Meta Built a New AI-Powered Ads Model for 5% Better Conversions](https://daily.dev/posts/how-meta-built-a-new-ai-powered-ads-model-for-5-better-conversions-plmz0bqth) · ByteByteGo · 0 upvotes · 0 comments - [Meta Details GEM Ads Model Using LLM-Scale Training, Hybrid Parallelism, and Knowledge Transfer](https://daily.dev/posts/meta-details-gem-ads-model-using-llm-scale-training-hybrid-parallelism-and-knowledge-transfer-21ryza9tt) · InfoQ · 0 upvotes · 0 comments - [Meta’s Generative Ads Model \(GEM\): The Central Brain Accelerating Ads Recommendation AI Innovation](https://daily.dev/posts/meta-s-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation-jduz74tbf) · Facebook Engineering · 1 upvotes · 0 comments --- Tags: [#nvidia](https://daily.dev/tags/nvidia), [#gpu](https://daily.dev/tags/gpu), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts), [#recommendation-systems](https://daily.dev/tags/recommendation-systems) [View this post on daily.dev](https://daily.dev/posts/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-model-system-co-design-q0bvljfan) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design","url":"https://daily.dev/posts/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-model-system-co-design-q0bvljfan","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-model-system-co-design-q0bvljfan"},"datePublished":"2026-10-28T16:04:57.399Z","dateModified":"2026-10-28T16:22:52.164Z","description":"Meta's next-generation Andromeda ads retrieval system now scores every eligible ad and surfaces top candidates at over one billion ads per second, a 100x...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/a348d419dbe06d4ba0b57f3725efe63d?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/a348d419dbe06d4ba0b57f3725efe63d?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Facebook Engineering\n","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Facebook Engineering\n","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/facebook","url":"https://daily.dev/sources/facebook_code"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-model-system-co-design-q0bvljfan","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"nvidia,gpu,mixture-of-experts,recommendation-systems","timeRequired":"PT12M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Facebook Engineering\n","item":"https://daily.dev/sources/facebook_code"},{"@type":"ListItem","position":3,"name":"Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/meta-andromeda-powering-the-next-generation-of-ads-retrieval-through-model-system-co-design-q0bvljfan#faq","mainEntity":[{"@type":"Question","name":"How much throughput improvement did Meta get from its 2026 Andromeda ads retrieval system compared to 2024?","acceptedAnswer":{"@type":"Answer","text":"The system achieved a 100x throughput lift over the 2024 version, now surfacing top ad candidates at over one billion ads per second on a single GPU server while scoring every eligible ad. The new mixture-of-experts gating architecture also reaches near ranking-level model quality using roughly 300x fewer inference-time FLOPs than the prior retrieval model. Engineers scaling ML inference systems can follow model-system co-design patterns like this one on daily.dev."}},{"@type":"Question","name":"How did Meta reduce network bottlenecks in its GPU-based ads retrieval pipeline?","acceptedAnswer":{"@type":"Answer","text":"Meta co-located index matching, scoring, and top-K selection onto a single GPU server instead of shipping candidate lists across separate hosts. The system now sends only a compact request payload of user features and a targeting query, cutting network traffic by about 100x and shifting the bottleneck back to GPU compute rather than network bandwidth. Developers designing distributed ML inference pipelines can track architecture trade-offs like these on daily.dev."}},{"@type":"Question","name":"What throughput gains did Meta and NVIDIA achieve by optimizing GPU kernels for ads retrieval?","acceptedAnswer":{"@type":"Answer","text":"The optimized interaction network (MoE gating) kernel improved 5.3x, from 2.4 billion to over 10 billion ads per second, reaching about 85% of peak HBM bandwidth. Index matching throughput improved over 2x to more than 100 billion ads per second. INT8 quantization added up to a further 2x gain over FP16 kernels with accuracy on par with 16-bit precision. Teams evaluating GPU kernel fusion and quantization trade-offs can find deep dives like this on daily.dev."}}]} ```