The development
What happened
Meta published a design for its next-generation Andromeda machine-learning system, which predicts user–ad relevance to select candidates before a later ranking stage. The company says its production pipeline scores every eligible ad at more than one billion ads per second, roughly 100 times the throughput of the 2024 system. These are Meta’s measurements, not independently verified results.
Key facts
The design combines a retrieval mixture-of-experts network with a fused GPU kernel, placing index matching, scoring and top-K selection on one GPU server.
Meta says the full retrieval model uses about 300 times fewer inference FLOPs than its 2024 model, while co-location cuts network traffic by about 100 times.
Meta reports that its interaction-network kernel processes over 10 billion ads per second on one NVIDIA GH200, reaching about 85% of peak HBM bandwidth.
A joint Meta-NVIDIA effort produced component speedups, but some full-pipeline gains remain benchmarks expected to be integrated into production later.
Meta reports a 2.4% increase in Facebook conversions since 2025; the supplied evidence does not establish an independent verification or complete cost comparison.
The context
Why it matters
Scoring all eligible ads during retrieval could change the trade-off between candidate selection quality and serving efficiency. The practical benefit remains uncertain without independent verification, a complete cost comparison or production results for all benchmarked optimizations.