SYSTEM Cited by 1 source
Meta Andromeda (Ads Retrieval Model)¶
Definition¶
Andromeda is Meta Ads's "next-gen personalized ads retrieval engine" introduced in 2024-12-02 Meta Production Engineering blog post. On the wiki it is referenced as the headline production beneficiary of KernelEvolve-generated kernels — KernelEvolve delivers >60% inference throughput improvement on Andromeda on NVIDIA GPUs over a baseline already optimized with torch.compile + vendor libraries (cuBLAS / cuDNN) (Source: sources/2026-04-02-meta-kernelevolve-how-metas-ranking-engineer-agent-optimizes-ai-infrastructure).
KernelEvolve impact — the baseline matters¶
The 60% number is meaningful precisely because the baseline is not naive code:
"On NVIDIA GPUs, it delivered more than 60% inference throughput improvement over a model with highly optimized kernels including torch.compile and vendor libraries — performance gains that directly translate to serving capacity and infrastructure efficiency."
KernelEvolve exceeds hand-tuned + compiler-tuned + vendor-library code, not just hand-written kernels — the delta represents a structural advantage of search-based + structured-diagnostic-feedback kernel optimization over compiler-heuristic + vendor-library-boilerplate.
Role in the Ads stack¶
Andromeda is the retrieval stage of Meta Ads's recommendation funnel; downstream stages include ranking models (the Meta Adaptive Ranking Model is the current LLM-scale incumbent). Together they form the retrieve-then-rank pipeline that picks which ads to serve on Facebook / Instagram / Meta properties.
Seen in¶
- Meta KernelEvolve (2026-04-02, canonical for KernelEvolve-impact axis). Named as the NVIDIA-GPU inference throughput headline beneficiary. (Source: sources/2026-04-02-meta-kernelevolve-how-metas-ranking-engineer-agent-optimizes-ai-infrastructure)
Caveats¶
The 2024-12-02 Andromeda Production-Engineering post is not ingested on this wiki. Architectural details of the model itself (candidate size, embedding shape, retrieval algorithm, training cadence) are not covered in the KernelEvolve post. The 60% inference throughput improvement is framed as applying to the Andromeda model overall; the specific kernels driving it are not itemized.
Related¶
- companies/meta
- systems/kernelevolve — the agentic system that generated the kernels producing the 60% improvement.
- systems/meta-adaptive-ranking-model — the downstream ranking stage in the Ads funnel.
- systems/meta-mtia — on the MTIA side, KernelEvolve-generated kernels deliver >25% training throughput on an (unnamed) ads model (may or may not be Andromeda).