JAXBench Released for Autonomous TPU Kernel Optimization

Arya Tschand, Charles Hong, Julian Walker, Nina Cai, Shangkun Wang, Suvinay Subramanian, Sundar Dev, Vijay Janapa Reddi, Amir Yazdanbakhsh, Sethu Sankaran· July 24, 2026 View original

Summary

Researchers have introduced JAXBench, a new benchmark suite for optimizing AI-generated kernels on Google Cloud TPUs, featuring 50 JAX workloads from production ML models and KernelBench. The benchmark evaluates autonomous optimization methods, showing that target-specific context and search structure significantly improve performance over XLA.

A new benchmark suite, JAXBench, has been released to drive progress in autonomous optimization of TPU kernels, similar to existing GPU benchmarks. This suite includes 50 JAX workloads, comprising 17 production machine learning operators from models like Llama-3.1 and AlphaFold2, and 33 translated operators from KernelBench, all optimized for high TPU v6e MXU utilization. The benchmark evaluates AI-generated kernel optimization methods, revealing that providing curated TPU documentation as context dramatically improves correctness from 5.8% to 37.3%, leading to a 1.28x geomean speedup. Furthermore, advanced search structures, such as Autocomp's beam-search pipeline, achieve a 1.36x geomean speedup over XLA, and up to 1.60x over XLA for hand-tuned kernels, though still trailing expert-level Pallas kernels in some specialized cases. The release of JAXBench aims to foster open-source contributions to high-quality TPU kernel optimization.

Why it matters

For professionals working with TPUs, this benchmark offers a standardized way to evaluate and improve the efficiency of AI-generated kernel optimizations, potentially leading to significant cost and performance gains in ML inference and training.

How to implement this in your domain

  1. 1Explore JAXBench to benchmark custom or AI-generated TPU kernels for performance.
  2. 2Integrate JAXBench into CI/CD pipelines for continuous performance monitoring of TPU workloads.
  3. 3Leverage the insights from the research regarding context and search structure to guide AI agent development for kernel optimization.
  4. 4Contribute to the open-source JAXBench project to expand its utility and scope.

Who benefits

AI DevelopmentCloud ComputingHigh-Performance ComputingResearch & Academia

Key takeaways

  • JAXBench provides a crucial benchmark for autonomous TPU kernel optimization.
  • Contextual information significantly boosts AI agent performance in kernel generation.
  • Advanced search strategies yield substantial speedups over standard XLA compilation.
  • High-quality TPU kernel optimization remains challenging but is critical for efficiency.

Original post by Arya Tschand, Charles Hong, Julian Walker, Nina Cai, Shangkun Wang, Suvinay Subramanian, Sundar Dev, Vijay Janapa Reddi, Amir Yazdanbakhsh, Sethu Sankaran

"arXiv:2607.20466v1 Announce Type: new Abstract: Rigorous benchmarks have driven progress in autonomous GPU kernel performance optimization by establishing a shared target to hillclimb on, but no equivalent exists for TPUs. We present JAXBench, a TPU-native benchmark suite for AI-…"

View on X

Originally posted by Arya Tschand, Charles Hong, Julian Walker, Nina Cai, Shangkun Wang, Suvinay Subramanian, Sundar Dev, Vijay Janapa Reddi, Amir Yazdanbakhsh, Sethu Sankaran on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses