Google's Axion host processors, based on the Arm architecture, further expand their application within Google's TPU portfolio, delivering the high performance and energy efficiency required for demanding AI workloads.
Google Cloud is adopting Google Axion processors in one of its latest TPU systems. Google Axion is a custom CPU built on the Arm Neoverse® platform. With the launch of the TPU 8i Max, Google will further expand the role of Axion CPUs in handling host computing for its AI accelerators, a role previously fulfilled in the TPU 8t and TPU 8i.
Google Axion provides the CPU computing foundation, enabling Google's AI infrastructure to handle large-scale training, real-time inference, and increasingly demanding agent AI workloads. Google's expansion of Axion to the TPU 8i Max marks another significant milestone in Arm's ongoing empowerment of core systems for next-generation AI infrastructure.
Driving Overall AI System Performance Improvements
AI performance is not determined by accelerators alone. CPUs are responsible for data preparation, workload orchestration, and running the infrastructure software that ensures the entire system operates efficiently. As agents perform reasoning, use tools, and coordinate complex tasks, the demand for CPU computing foundations continues to grow.
With Axion, Google combines the performance and energy efficiency advantages of Neoverse with custom chips built for its own infrastructure, extending this approach to the TPU 8i Max. This further highlights the importance of co-designing CPUs and accelerators to drive AI scalability.
Axion CPUs, based on the Arm Neoverse platform, provide host computing power for the TPU 8t, TPU 8i, and TPU 8i Max.
Mark Lohmeyer, Vice President and General Manager of Google Cloud AI and Compute Infrastructure, stated: "Unlocking AI performance at scale requires co-designing CPUs and accelerators to enable seamless collaboration. Google's Axion processors, based on the Arm architecture, deliver the performance, energy efficiency, and flexibility required for high-demand TPU system host workloads. With the TPU 8i Max, we are further expanding this co-design foundation to help customers tackle their most complex AI application challenges."
ushering in a New Era of Inference
Google Cloud's continued investment reflects a broad industry trend: leveraging the synergy between CPUs and accelerators to usher in a new era of AI inference. In addition to providing host computing for accelerators, Google Axion also supports general-purpose cloud instances running agent runtimes and orchestration frameworks.
Earlier this year, Google Cloud launched the Google Axion N4A, offering up to twice the price-performance ratio compared to same-generation, comparable x86 virtual machines. By running the GKE Agent Sandbox on Axion N4A, customers can securely isolate agent AI workloads and improve price-performance by 30%. Building on this, the Agent Substrate on GKE also natively supports Google Axion processors, providing an open-source, pre-hardened agent execution runtime designed to support millions of sandboxes, achieving densities ten times that of standard container runtimes.
From powering Google's latest TPU systems to supporting the scaled operation of millions of AI agents, Arm is helping to build the computing foundation for the next generation of AI.
– End –