Google Cloud is expanding its offering with new Axion-based virtual machines (VMs) and Ironwood Tensor Processing Units (TPUs), providing users with significant advancements in performance and efficiency. These innovations are designed to cater to a range of workloads, from data processing to AI/ML tasks, while addressing both cost and sustainability goals.
The new Axion C4A and N4A VMs are powered by Google’s custom ARM-based Axion processors, built on the Arm Neoverse N3 core architecture. These VMs offer substantial improvements over traditional x86-based VMs, with the C4A series providing up to 65% better price-performance and 60% improved energy efficiency. The N4A series, currently in preview, promises up to a 105% increase in price performance and an 80% boost in performance-per-watt. These VMs support up to 64 vCPUs, 512GB DDR5 memory, and 50 Gbps networking, making them ideal for high-performance computing tasks.
An important aspect of these Axion-based VMs is their seamless integration with Databricks on Google Cloud and newly optimized Ubuntu images. This integration allows companies to leverage existing SQL, AI/ML, and ETL workloads without requiring code changes, making it possible for early adopters like Epsilon and Vimeo to report 20-25% runtime reductions and 10-15% cost savings, coupled with 30-60% performance improvements.
Simultaneously, Google’s latest Ironwood TPUs (version 7) aim to revolutionize AI processing. These TPUs are Google’s response to Nvidia’s Blackwell GPUs, optimized for TensorFlow workloads, and boast an impressive 4.6 petaFLOPS of FP8 compute with 192GB HBM3e memory. The Ironwood TPUs feature massive scalability with options to configure up to 9,216 chips into superpods, providing robust solutions for enterprise-level AI training and inference needs.
Major AI companies, including Anthropic, are already adopting this TPU infrastructure, with plans to deploy large numbers of TPUs to train next-generation models. This, combined with Google’s acquisition of $9.8 billion worth of TPUs from Broadcom, shows a strong commitment to providing cutting-edge AI solutions that challenge existing GPU-based paradigms.
In summary, Google’s latest releases — Axion-based VMs and Ironwood TPUs — provide enterprises with a powerful suite of tools to enhance performance, reduce costs, and pursue their sustainability commitments efficiently. Organizations can now choose from cost-optimized (N4A) to high-performance (C4A) Axion instances, and leverage the game-changing TPU v7 for their AI-centric functions.

Collections

Google Cloud introduces Axion-based ARM virtual machines and Ironwood TPU v7 processors for improved AI and data workloads. The C4A and N4A VM series deliver up to 105% better price-performance and 80% energy efficiency gains over x86 alternatives, supporting up to 64 vCPUs and 512GB memory. Ironwood TPUs offer 4.6 petaFLOPS compute with 192GB HBM3e memory, scalable to 9,216-chip superpods for enterprise AI training. Early adopters report 20-25% runtime reductions and significant cost savings, with seamless integration into existing Databricks and Ubuntu environments requiring no code changes.

Google Cloud Introduces Axion-Based N4A VMs in Preview with New Ironwood TPUs