Meta Andromeda next generation: one billion ads per second, 100x throughput

What ads retrieval does

Retrieval is the first stage of Meta's ads recommendation pipeline: it narrows a vast pool of eligible ads down to a small set of strong candidates in milliseconds. A later, more expensive ranking stage then scores those survivors in detail. Meta says applying ranking-stage model quality at retrieval scale had historically been too costly.

What the new generation brings

Meta says the new Andromeda co-designs the model, system and GPU hardware to bring near ranking-level model performance to retrieval scale. In production, the system scores every eligible ad at that quality and surfaces the top candidates at over one billion ads per second, which Meta describes as a 100x lift over the 2024 Andromeda system.

Why the architecture changed

The 2024 deployment ran on the NVIDIA GH200 Grace Hopper Superchip and grew model capacity by four orders of magnitude. As generative AI creative tools pushed the ads candidate pool to hundreds of millions of new ads, Meta needed a fundamental architectural change to keep improving performance at a sustainable compute cost.

The new bottleneck: memory and network

At billions of ads per second, Meta says memory bandwidth becomes the bottleneck. Intermediate tensors between GPU kernels account for terabytes per second of memory traffic, and shipping ads between local CPU hosts and remote accelerators saturates the network link. The fix is a new system design that unifies fragmented, distributed workloads into a single GPU pipeline.

What it means for advertisers

Meta says advertisers get better matching across a larger candidate pool within the same latency budget. The company reports that Andromeda has delivered a 2.4% increase in conversions on Facebook since 2025.