# Meta Andromeda: Scaling Ads Retrieval Through Model-System Co-Design Meta achieves over one billion ads per second throughput using unified GPU-accelerated pipeline architectures. - **Source:** [Engineering at Meta](https://brifing.si/source/28383801-e365-4f15-8f20-49d8c0635946) - **Published:** 2026-10-28 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/4dcbb57e-43c3-490d-ba25-1b5b98e0eaae ## Evolution of Meta Andromeda Meta Andromeda serves as the primary machine learning system for personalized ads retrieval, acting as the initial stage of the recommendation pipeline. Since 2025, this system has delivered a 2.4% increase in conversions on Facebook. It effectively narrows down a vast pool of eligible advertisements into a set of high-quality candidates. ## Breaking Throughput Benchmarks The latest iteration of Meta Andromeda achieves a throughput exceeding one billion ads per second. This represents a 100x performance lift compared to the 2024 system deployment. By scoring every eligible ad with near ranking-level precision, the system maximizes advertiser value while maintaining strict latency budgets for real-time ad delivery across platforms. ## The Memory Bandwidth Bottleneck At a scale of billions of ads per second, memory bandwidth becomes the primary system constraint. Moving ad tensors through the memory hierarchy creates terabytes of traffic per second. Previous multi-kernel pipelines suffered from excessive data movement, which saturated network links and wasted GPU compute cycles on redundant intermediate tensor materialization tasks. ## Retrieval Mixture-of-Experts Architecture The new architecture integrates LLM-inspired mixture-of-experts into the retrieval stage. By transitioning from single user towers to ensembles of specialized experts, the model captures diverse, time-evolving user interests. This design achieves near ranking-level performance while requiring approximately 300x fewer inference-time FLOPs than the 2024 model, significantly improving overall system efficiency and capacity. ## Fused GPU Kernel Innovation To eliminate memory bottlenecks, engineers developed a fused GPU kernel that combines model execution and system logic. By performing all operations in a single pass, intermediate results remain in GPU registers and shared memory. This approach prevents materialization to high bandwidth memory, allowing the system to reach 10 billion ads per second. ## Co-Location Strategy Benefits The system architecture was collapsed to co-locate index matching, scoring, and top-K selection on a single GPU server. This strategy reduces network traffic by approximately 100x compared to previous distributed approaches. By sending only compact request payloads, the system shifts from being network-bound to being effectively compute-bound on the local GPU hardware. ## Optimizing Index Matching Logic Moving complex advertiser-defined targeting logic onto the GPU required a complete redesign of the index structure. The new structure supports GPU-parallel execution while maintaining full fidelity for real-time updates. This optimization allows index matching to reach speeds over 100 billion ads per second, ensuring it remains well ahead of the scoring network. ## Joint Collaboration with NVIDIA Meta and NVIDIA collaborated in early 2026 to push retrieval pipelines toward hardware limits. Each component of the pipeline achieved between 2x and 5x speedups over the baseline. This partnership validated that additional model complexity, such as MoE gating, is essentially free when executed on specialized Tensor Core hardware at high throughput. ## Quantization for Efficiency The team implemented INT8 quantization to further enhance throughput beyond the optimized FP16 kernels. This technique provided up to a 2x increase in processing speed without sacrificing model accuracy. By maintaining 16-bit precision performance levels, the system maximizes the utility of the GH200 Grace Hopper Superchip for large-scale ad candidate scoring. ## Future Platform Scaling Plans Building on successful Hopper results, the team is extending co-design efforts to the Grace Blackwell platform. Current research focuses on optimizing CPU-to-GPU compute ratios to better suit Andromeda-like workloads. These platform design optimizations aim to determine how sensitive end-to-end application performance is to varying hardware configurations in future large-scale deployments. ## Impact on Advertiser Value Ads retrieval provides an outsized return on investment for incremental quality improvements. By enabling richer feature interactions and deeper architectures without proportional cost increases, Meta allows advertisers to reach audiences more effectively. The system ensures that even as the candidate pool grows, the relevance of served ads remains high for users. ## Strategic Significance of Co-Design The Meta Andromeda project demonstrates that hardware-software co-design is essential for modern AI-driven recommendation systems. By unifying model architecture with system execution, Meta has established a new frontier for retrieval efficiency. This approach allows for sustainable growth in model capacity, ensuring the platform remains performant as ad candidate pools continue to expand. ## Sources covering this story - https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._