Google Updated XProf With Custom Kernel Profiling

The new suite enables cycle-level performance tracking for custom MLIR kernels on TPU v7 Ironwood hardware.

Updated on Sept. 23, 2026 in Semiconductors

Isometric editorial illustration of a black TPU processor chip with copper circuitry lines, representing granular hardware-level kernel performance analysis. — Google has launched a new kernel profiling suite for its TPU v7 hardware, enabling engineers to perform cycle-level debugging of custom machine learning code. AI Illustration. Upload story photo >

Live Poll

Is it worth the effort to manually optimize code using low-level kernel profiling tools?

Google has released a Kernel Profiling suite for its XProf tool, designed to provide granular cycle-level visibility into custom machine learning kernels. This update introduces sub-microsecond capture capabilities for specialized workloads running on TPU v7 Ironwood chips.

Why it matters

Custom kernels built via Pallas, Mosaic, or Triton often bypass standard XLA compiler optimizations, rendering static cost models inaccurate. This tool provides the necessary visibility to debug and tune these kernels for improved hardware utilization.

The suite exposes 16,000 raw performance counters and supports configuring up to 28 counters across four SparseCores. It provides sub-microsecond capture, improving upon the 1µs resolution floor of standard host-level timers.

The players

Google

A global technology company developing large-scale distributed systems and proprietary hardware architectures like the Tensor Processing Unit.

OpenXLA

An open-source machine learning compiler ecosystem focused on accelerating deep learning models across diverse hardware backends.

The details

The profiling suite integrates directly with the OpenXLA project and JAX to track machine instructions, including MXU (Matrix Multiplication Unit) and ALU (Arithmetic Logic Unit) operations, per clock cycle. By passing specific flags for compiler inspection, developers can generate a Custom Call Text panel that displays the lowered MLIR (Multi-Level Intermediate Representation) code. This allows engineers to observe how custom calls map to hardware execution on TPU v7 chips, bypassing the limitations of traditional black-box performance analysis.

September 23, 2026: Google announced the new XProf Kernel Profiling suite.

The Tech Race

This release marks an attempt to bridge the gap between high-level JAX abstractions and low-level hardware performance tuning on custom silicon. It directly addresses the visibility limitations inherent in competing JIT-compiled ML workflows that skip standard optimization passes.

Developers working with JAX and Pallas can now inspect lowered machine instructions to diagnose performance bottlenecks that were previously opaque. Implementation requires updating to the latest XProf tooling and configuring specific compiler flags to enable the Custom Call Text panel.

The takeaway

As machine learning kernels become more specialized, hardware-specific profiling is shifting from an optional optimization to a mandatory engineering step. Watch for future updates to the OpenXLA project that may automate the tuning process based on these new granular performance metrics.