Meta says redesigned Andromeda scores over one billion ads per second

Meta says its redesigned Andromeda advertising system processes over one billion ads per second on a single GPU server in production, a 100x throughput increase over its 2024 system. The system selects candidates for a later, more detailed ranking stage while scoring every eligible ad within the same latency budget. The redesign routes ads among specialized models and moves targeting checks, scoring and final candidate selection onto the same GPU. Keeping intermediate results there reduces memory traffic, while co-location cuts network traffic by approximately 100x. Further optimizations developed with NVIDIA benchmark at multiple times current production throughput, but those additional gains are still being integrated into production.