Meta uses idle inference capacity for ads recommendations
Meta combines locally idle pockets across regions and accelerator types into a shared, opportunistic inference pool, even while the fleet is constrained overall. Its ads-ranking system uses two elasticity levers: Rank More Ads evaluates additional candidates, while Serve Better Models changes the share of requests sent to a more powerful model.
Event / data period:
Published: · Updated:
Rank More Ads needs surplus CPU and accelerator capacity for additional candidates. Serve Better Models keeps candidate volume fixed and routes selected requests to the stronger model, allowing it to use accelerator-only surplus. The two paths can consume different mixes of spare resources.
Meta isolates baseline serving from optional work physically or logically. Central control loops adjust over minutes, while server- and client-side load shedding can drop lower-value optional work within milliseconds if capacity is reclaimed.
The request router selects which requests receive the stronger model through an explore-and-exploit process that periodically measures incremental value; when capacity shrinks, those requests can fall back to the baseline model. Meta’s post details architecture and controls but gives no quantified uplift in ad performance.
Data
Meta uses idle inference capacity for ads recommendations
- · Occurred
Meta uses idle inference capacity for ads recommendations
Meta combines locally idle pockets across regions and accelerator types into a shared, opportunistic inference pool, even while the fleet is constrained overall. Its ads-ranking system uses two elasticity levers: Rank More Ads evaluates additional candidates, while Serve Better Models changes the share of requests sent to a more powerful model.
Methodology
This event record summarizes the dated statement in the cited primary source. Product or test claims from the source are not presented as independently verified results.
Updated:
Subjects and places
Subject proposals
- Science and technologySuggested
- Preserved editorial category/topic; broad candidate onlyengineering.fb.com ↗