Google Adds Cycle-Level Kernel Profiling to XProf

Google has added a Kernel Profiling suite to XProf, its open-source profiler for TPU workloads. Developers can now see cycle-level details in custom Pallas kernels that previously appeared as opaque blocks in traces. In a demo on TPU v7 (Ironwood), triple buffering cut kernel time from 125.5µs to 88µs, about 30%.

• XProf exposes cycle-level metrics for Pallas, Mosaic and Triton kernels on TPU v7
• Counter sampling has a 1µs resolution floor; an external trigger mode removes it
• Up to 28 counters per core across up to four SparseCores (4x28 matrix)
• In the demo, triple buffering cut matmul kernel time from 125.5µs to 88µs