Meta has rolled out Zoomer, an automated platform aimed at optimizing AI performance through debugging and analysis. It operates across all AI workloads, including ads ranking and generative AI features, providing insights that help improve efficiency and reduce energy consumption. By addressing performance issues, Zoomer has cut training times and boosted query per second (QPS) across Meta’s AI infrastructure. Zoomer’s architecture includes three layers: the infrastructure and platform layer for scalability, the analytics engine for deep performance insights, and a user-friendly visualization layer. The platform collects a wide array of data, from GPU utilization and memory usage to execution traces and application-level annotations, enabling comprehensive performance analysis. Profiling can be automatic or on-demand, ensuring timely insights based on workload type. The automated analysis pipeline identifies various performance issues, such as straggler detection and bottleneck analysis, streamlining the optimization process. Results are presented through multiple interfaces, including interactive visualizations and detailed dashboards, making it easier for teams to address inefficiencies. By focusing on maximizing GPU utilization and minimizing waste, Zoomer plays a crucial role in meta's strategy to support large-scale AI workloads while reducing environmental impact.