In a recent research report, Google provides insight into one of the largest technological operations within the company in recent years. It concerns the migration of its entire software ecosystem from x86 processors to Arm.
The research, entitled Instruction Set Migration at Warehouse Scale, describes how this transition took place and the role artificial intelligence played in automating the process.
The report makes it clear that the difficulty of such migrations no longer lies in translating machine language or rewriting source code. Whereas in the past, a lot of attention was paid to binary translation, the reality today is that modern compilers and open-source ecosystems make this step largely unnecessary. The real challenge lies in the enormous number of small, often trivial adjustments scattered across millions of lines of code and thousands of configuration files.
Google analyzed a total of 38,000 code commits that were part of the migration. This revealed that only a fraction related to actual code translation, while the majority consisted of adjustments to build systems, test configurations, and infrastructure settings. The researchers emphasize that these tasks are simple in themselves, but their scale and distribution across the entire code base pose a huge coordination challenge.
Migration largely without developers
To address this problem, Google deployed its internal automation system, Rosie. This enabled the company to automatically generate, test, and submit thousands of small changes to the appropriate teams. Another system, CHAMP, then checked whether the new Arm versions of software ran as stably and reliably as their x86 counterparts. Thanks to this approach, most of the migration took place without the direct intervention of developers.
Google also experimented with the use of generative AI. An internally developed agent, CogniPort, was tasked with independently repairing failed builds and tests. This agent worked in a cyclical process of analysis, adjustment, and retesting. In about 30% of cases, the system managed to repair errors completely automatically. Although this percentage is still limited, according to the researchers, it shows that artificial intelligence has a promising role to play in future migrations.
According to The Register, Google has now migrated approximately 30,000 production packages to Arm, with another 70,000 applications in the queue. Services such as YouTube, Gmail, and BigQuery already run on both x86 and Google’s own Axion chips, an internally developed Arm variant. These machines are said to perform up to 65 percent better in terms of price-performance ratio and be 60 percent more energy-efficient than their x86 counterparts. The switch enables Google’s Borg cluster system to flexibly distribute workloads across different architectures, which should further increase efficiency in data centers.
Google’s transition to Arm proves that even a code base of billions of lines can be transformed, provided that automation and artificial intelligence are central to the process.