Google Added Kernel Profiling Suite to XProf

The new tool provides cycle-level performance data for custom TPU workloads.

Updated on Sept. 23, 2026 in Semiconductors

Isometric editorial illustration of a silicon chip on a plinth surrounded by radiating beams, representing high-performance machine learning data analysis. — Google has introduced the XProf Kernel Profiling suite, allowing engineers to track custom machine learning workloads on TPU v7 Ironwood hardware for improved efficiency. AI Illustration. Upload story photo >

Live Poll

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

Google has launched a new XProf Kernel Profiling suite designed to track custom Pallas, Mosaic, and Triton kernels on TPU v7 Ironwood chips. This update offers engineers granular visibility into hardware operations that previously bypassed standard cost models.

Why it matters

Custom-built kernels often distort traditional static cost models, making it difficult for developers to optimize machine learning performance. This suite allows for precise instruction tracking to improve execution efficiency in complex AI workloads.

The suite tracks machine instructions across four SparseCores, supporting the configuration of up to 28 counters per core. The Perf Counters View provides access to over 16,000 raw metrics.

The players

Google

Google is a major technology company that develops the XProf suite and TPU hardware architectures.

OpenXLA

OpenXLA is an open-source project focused on creating a domain-specific compiler for machine learning.

JAX

JAX is a high-performance numerical computing library maintained by Google for machine learning research.

The details

The profiling suite integrates with OpenXLA and JAX to display lowered MLIR via a new Custom Call Text panel. An external event-triggered mode now enables sub-microsecond capture, allowing developers to monitor specific TPU trace instructions during runtime.

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

Roadmap

This development aligns with the shift toward specialized hardware-aware software engineering in the semiconductor industry. As AI models outgrow general-purpose cost models, tooling like XProf is becoming essential for maintaining competitive advantage in custom silicon performance.

Developers working on custom ML workloads can now achieve significant performance gains by identifying bottlenecks at the instruction level. This reduces the time required for model training and deployment processes.

The takeaway

Precision tracking of kernel execution allows developers to maximize the throughput of specialized hardware. Implementing these tools early can lead to measurable efficiency improvements in complex machine learning pipelines.