Google Adds Cycle-Level Kernel Profiling to XProf

Google added a Kernel Profiling suite to XProf, its open-source profiler for TPU workloads, letting developers see cycle-level detail in custom Pallas kernels that previously appeared as opaque blocks. On TPU v7 (Ironwood), XProf samples hardware performance counters at runtime. In a tiled matmul demo, triple buffering cut kernel time from 125.5 microseconds to 88 microseconds, roughly 30 percent.