Photo by Dima Solomin on Unsplash
A 2.4% lift in real time
When a major consumer‑goods brand examined its latest Facebook campaign, the dashboard showed a 2.4% rise in conversions attributed to the newly deployed Andromeda engine. The boost came without any change to creative spend, simply by switching the retrieval layer that decides which ads enter the auction. Andromeda is Meta’s machine‑learning system for personalized ads retrieval, the first stage that narrows the pool of eligible ads before the ranking model takes over.
How retrieval reshapes ad selection
Andromeda works at the very first stage of the ads pipeline, scoring every eligible ad with near ranking‑level quality. It then narrows a pool of tens of millions of ads to a handful of strong candidates in milliseconds, effectively turning a massive catalog into a curated shortlist for each user. The system can surface top candidates at a rate of over one billion ads per second, a benchmark that dwarfs prior retrieval engines. By moving ranking‑grade modeling to the retrieval stage, advertisers receive better matches without any extra latency, letting the auction choose from a richer set of options.
From 2024 launch to hardware co‑design
The engine debuted in 2024 on NVIDIA’s GH200 Grace Hopper Superchip, pairing a custom model with hardware designed for massive parallelism. A co‑design effort that aligned model architecture, system software and GPU capabilities produced a 100‑fold lift in throughput compared with the earlier 2024 system. The 2025 update introduced the 2.4% conversion lift now being reported. Meta’s earnings release scheduled for late January 2026 will likely include the financial impact of these ad efficiency gains.
The engineers, chips and advertisers
The research paper lists a team of engineers and scientists from Meta’s AI division, including Haiyu Lu, Shuo Gu, Yang Yang (Tony) and several others who contributed to the system’s design. NVIDIA supplied the GH200 Grace Hopper Superchip that powers the retrieval stage, enabling the system to evaluate billions of candidates in real time. Because the retrieval stage processes every eligible ad, any advertiser running a Facebook campaign feeds their creative assets into the same pipeline, trusting the engine to surface the most relevant message for each user.
What the boost means for you
At a throughput of over one billion ads per second, a 2.4% improvement touches a massive volume of impressions, turning a modest percentage gain into a sizable impact on overall performance. The result shows that moving sophisticated modeling earlier in the pipeline can deliver measurable revenue without altering creative budgets. For marketers, the lesson is clear: the technology that decides which ad to show can be as important as the ad itself, and advances in retrieval efficiency are already reshaping the economics of digital advertising.
Sources
- Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design - Engineering at Meta — “Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook.”
- Meta sets Q4 and 2025 results for Jan. 28 | META Stock News — “Meta Platforms, Inc. (NASDAQ: META ) announced today that the company's fourth quarter and full year 2025 financial results will be released after market close on Wednesday, January 28th, 2026.”
Top comments (0)