Arm has introduced its latest Armv9 Edge AI Platform, designed to propel IoT applications by combining the high efficiency of the Cortex-A320 processor and the cutting-edge Ethos-U85 neural processing unit (NPU). Paul Williamson, senior vice president and general manager of Arm IoT business, said that this platform will bring a more efficient computing experience to the IoT ecosystem, allowing different workloads to run in the most appropriate place and at the most appropriate time.
Cortex-A320: Ultra-Efficient Armv9 CPU Optimized for IoT
The Cortex-A320 is Arm’s first ultra-efficient CPU based on the Armv9 architecture, delivering a 10× increase in ML performance and a 30% improvement in scalar performance compared to its predecessor, the Cortex-A35. These advancements are largely attributed to the integration of Scalable Vector Extension 2 (SVE2), which enhances vector processing and AI workload execution.
Designed for heterogeneous compute environments, the Cortex-A320 operates seamlessly with the Ethos-U85 NPU, dynamically distributing workloads to minimize latency and optimize power consumption.
“Our V9 Edge AI platform couples the ultra-efficient Cortex-A320 with the high-performance Ethos-U85 neural processor, delivering a truly heterogeneous edge AI solution,” said Williamson. “It enables the IoT ecosystem to execute workloads efficiently—placing them in the right location at the right time.”
To support scalability, the Cortex-A320 features an enhanced memory architecture capable of handling larger AI models with multi-tiered memory access. It can also be configured in clusters of up to four cores, providing flexibility for different application needs.
Ethos-U85: Arm’s Most Advanced NPU for AI Acceleration
The Ethos-U85 is Arm’s most advanced NPU, delivering up to 4 tera operations per second (TOPs) at 1 GHz while improving power efficiency by 20% over its predecessors, the Ethos-U65 and U55.
A major breakthrough in this NPU is its native support for transformer-based AI models, which are becoming critical in computer vision and natural language processing applications.
Additionally, the direct integration of the Cortex-A320 and Ethos-U85 eliminates the need for a separate Cortex-M controller, simplifying the system architecture and reducing both cost and complexity.
Arm Kleidi: Unlocking AI Performance Without Added Complexity
A key differentiator of the Armv9 Edge AI Platform is Arm Kleidi, a suite of compute libraries that automatically optimizes AI workloads on Arm CPUs, eliminating the need for extensive manual tuning by developers.
Initially introduced for Arm’s client markets, Kleidi has now been extended to IoT applications, enabling developers to achieve up to 70% performance gains on ML workloads running on the Cortex-A320.
“Kleidi AI provides developers with frictionless access to the best performance possible on an Arm CPU,” Williamson explained.
The Kleidi suite includes:
● KleidiAI – Optimizes ML computations.
● KleidiCV – Enhances computer vision workloads.
These libraries integrate with leading AI frameworks such as Llama.cpp, ExecuTorch, and LiteRT, and leverage Neon and SVE2 optimizations to accelerate vector processing.
Industry Adoption and Future Implications
Arm’s new Edge AI platform is already generating strong interest from industry leaders. Companies such as AWS, Siemens, Renesas, and Advantech are exploring its potential applications:
● AWS plans to integrate the Cortex-A320 with its Nucleus Lite runtime for resource-constrained edge devices.
● Siemens views the platform as a foundation for secure and high-performance AI in industrial automation.
Williamson emphasized the broader impact of AI at the edge, stating:
“AI at the edge is not just about improving commercial applications—it has the potential to address global challenges, from reducing the climate impact of human activity to enhancing biodiversity monitoring and revolutionizing healthcare.”
As edge AI continues to evolve, Arm’s Armv9 Edge AI Platform is set to play a crucial role in enabling next-generation IoT solutions, pushing the boundaries of efficiency, scalability, and AI-driven intelligence at the edge.
