How Google Is Moving 30,000 Production Packages to Arm with AI‑Powered CogniPort

Google announced it is porting roughly 30,000 production software packages to the Arm architecture, aiming to run them on its own Axion chips as well as x86 processors.

Last week the company published a preprint titled “Instruction‑Set Migration at Repository Scale” ( https://arxiv.org/pdf/2510.14928 ) describing the effort.

The paper notes that major Google services such as YouTube, Gmail and BigQuery already run on both x86 and Axion Arm CPUs, with about 30,000 applications continuing to operate.

Engineers Parthasarathy Ranganathan and Wolff Dobson explain that the migration began with the hypothesis of studying architectural differences—floating‑point drift, concurrency, platform‑specific operators, and performance.

Early work focused on key services like F1, Spanner and Bigtable, using standard software practices, weekly meetings and dedicated engineers. They found fewer issues than expected because modern compilers and cleanup tools mitigated many problems.

Fix tests that failed due to over‑fitting to existing x86 servers.

Update complex build and release systems, especially for the oldest and highest‑traffic services.

Resolve rollout issues in production configurations.

Ensure stability of critical systems.

Google is also developing an AI‑driven tool called “CogniPort” to automate the migration. CogniPort intervenes when an Arm library, binary, or test fails to build or passes with errors, automatically generating fixes; its blueprint editor has already produced migration commits that cannot be changed manually.

Initial experiments show CogniPort succeeds about 30 % of the time, performing best on test‑fixes, platform‑specific conditions and data‑representation fixes. Google still has roughly 70,000 packages left to port.

The ultimate goal is to enable Borg, Google’s internal cluster manager built on Kubernetes, to schedule workloads efficiently on Arm servers, which Google claims are 65 % cheaper and 60 % more energy‑efficient than comparable x86 instances.