Meta rebuilt ads retrieval around a mixture-of-experts model and a fused GPU kernel, and says it now scores over one billion ads a second at roughly 300x fewer inference FLOPs
AspireAI community · Meta Engineering
Ads recommendation has two stages: retrieval, which sweeps a huge pool of eligible ads down to a shortlist in milliseconds, and ranking, which then scores that shortlist properly. Ranking is the more accurate stage and has always been too expensive to run at retrieval scale. Meta's new Andromeda generation is an attempt to close that gap by moving the model, the serving system and the GPU kernel into a single co-designed pipeline.
Three changes do the work. A retrieval mixture-of-experts model routes each candidate to the most relevant user experts, which Meta says reaches near-ranking-level quality at roughly 300 times fewer inference FLOPs than the 2024 retrieval model. A fused GPU kernel runs the whole per user-ad computation, model and system logic together, in one pass instead of many small kernels, which keeps intermediate results in registers rather than pushing them through memory. And per-user request work is co-located with the GPU scoring, cutting network traffic by about 100 times. End to end the pipeline scores every eligible ad across Facebook and Instagram at over one billion ads per second on a single GPU server, with the isolated model kernel reported above ten billion per second on one GH200.
The limits are worth stating plainly. These are Meta's own measurements of an internal system, and the 2.4% conversion lift is credited to Andromeda since 2025 overall rather than to this iteration. The ten-billion figure is the isolated kernel, not the shipping pipeline, and the Grace Blackwell work is still an evaluation in progress.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.