Explore 3 capabilities: Continuous Profiling Snapshots, the Utilizatio

Google Open Source is Google's initiative for promoting open-source software development and collaboration, offering tools, resources, and best practices for open-source projects. Developers can learn about Google's open-source projects, contribution guidelines, and community engagement to participate in and contribute to open-source projects effectively.

Google Open Source Blog

Google's XProf profiler introduces three advanced TPU optimization capabilities: Continuous Profiling Snapshots (an always-on flight recorder using a ~2GB circular buffer to capture the last ~90 seconds of performance data with ~7µs overhead), the Utilization Viewer (translates raw hardware counters into readable utilization percentages for TensorCore, SparseCore, and HBM), and LLO Bundle Visualization (exposes machine-level instruction scheduling data for kernel developers using Pallas, enabling cycle-level bottleneck analysis in the MXU pipeline). These tools are now available in the open-source XProf repository under OpenXLA.

Advanced TPU optimization with XProf: Continuous profiling, utilization insights, and LLO bundles

Continuous Profiling Snapshots: The "Flight Recorder" for ML

Visualizing Hardware Efficiency with the Utilization Viewer

Inspecting the Metal: Low-Level Operations (LLO) Bundles