Google porting all internal workloads to Arm, with AI helper

Steven

Google porting all internal workloads to Arm, with AI helper

Summary

Google has ported roughly 30,000 production packages to Arm (its Axion CPUs) and intends to convert all internal workloads so services can run on both Axion and x86. The company published a preprint, “Instruction Set Migration at Warehouse Scale”, and a blog post explaining the process. Engineers found fewer low-level ISA issues than expected — modern compilers and sanitizers removed many surprises — and most work ended up on brittle tests, legacy build/release systems and production rollout configurations.

To scale the effort Google extended automation and built a GenAI agent called CogniPort to operate on build and test errors. CogniPort can generate migration commits and automatically fix certain classes of failures; it succeeds about 30% of the time under specific conditions. Around 70,000 packages remain to be migrated. The goal is to enable Borg (the basis for Kubernetes) to schedule workloads across architectures and exploit Axion’s claimed up-to-65% better price-performance and up-to-60% energy efficiency compared with x86.

Key Points

  • About 30,000 production packages have already been ported to Arm; roughly 70,000 remain.
  • Major services — including YouTube, Gmail and BigQuery — already run on both x86 and Google’s Axion Arm CPUs.
  • Fewer architecture-specific bugs were found than expected; most effort went into fixing tests, updating build/release pipelines and smoothing rollouts.
  • Google created CogniPort, a GenAI-based agent that works on build/test errors and automatically generates fixes; it has ~30% success in targeted scenarios.
  • Objective is multi-architecture scheduling in Borg/Kubernetes to capture Axion’s claimed cost and energy advantages, potentially reducing future x86 demand.

Context and relevance

This is a strategic, large-scale migration with industry-wide implications. It confirms continued momentum towards Arm in datacentres (alongside offerings like AWS Graviton), and it demonstrates a practical use of GenAI to automate repetitive engineering fixes at scale. The effort could shift server purchasing, cloud economics and how operators plan multi-arch deployments.

Why should I read this?

Short and blunt: it’s big. Google is not just testing Arm — it’s actively rewiring its fleet and using AI to do the heavy lifting. If you care about cloud costs, server architecture, or how AI is changing engineering workflows, this saves you time and tells you what to watch for next.

Author style

Punchy: this isn’t a small optimisation — it’s a strategic migration that could reshape server economics and the future role of x86. Read the details if you want the numbers and technical caveats; skim if you only want the headlines.

Source

Source: https://www.theregister.com/2025/10/22/google_multi_arch_x86_arm_port/