Arm announced an Armv9-based AI platform that combines the new Arm Cortex-A320 CPU with the Arm Ethos-U85 NPU, a major AI accelerator for edge AI. It is optimized for IoT and is capable of running AI models with over 1 billion parameters on the device. Arm President and CEO Takayuki Yokoyama said, “What we are announcing this time is not cloud-side AI, but edge-side AI. As AI has developed to this extent, it is becoming very difficult to run all AI on the cloud side. As the AI revolution progresses rapidly, it is becoming more important than ever to run the right AI workload in the right place to make further progress.” The Armv9 edge AI platform announced is characterized by its concentration of efficiency, security, and intelligence on the edge side. Compared to the Cortex-M85-based platform announced in 2024, machine learning (ML) performance is said to be eight times better. Director of the Applied Technology Department, Masashi Nakajima, emphasized, “Transformer-based generative AI has appeared, and the trend of applying it is accelerating.
The Cortex-A320 announced this time is the most compact CPU that uses the Armv9 architecture. Although it is small in size, by going 64-bit, it will be possible to use very large parameter models.” The Cortex-A320 utilizes features of the Armv9 architecture such as “SVE (Scalable Vector Extension) 2” to improve ML performance, and compared to the previous generation “Cortex-A35”, it can improve ML performance by 10 times and scalar performance by 30%. In addition, the pipeline has been revised to create a highly energy-efficient processor. Compared to the higher-end model “Cortex-A520”, it can perform the same processing with about 50% of the power consumption. “The biggest advantage of using Armv9 for IoT devices is the enhanced security and AI computing performance.
The architecture has been enhanced in various ways, and features such as Pointer Authentication (PAC) and Branch Target Identification (BTI) are moving in the direction of eliminating software vulnerabilities,” said Nakajima. In the future, the plan is to expand use cases at the edge by running large language models (LLM) and small language models (SLM) tuned for agent-based AI applications. “I think that edge AI is the most important thing in the world of IoT. Until a few years ago, there were many relatively simple functions such as filtering to remove noise and anomaly detection, but now, workloads on the edge side are becoming more complex to meet the demands of advanced use cases. To execute this, I think we need higher performance and better efficiency than now. There is also a high demand from fields such as factory automation (FA), smart cities, and smart homes,” said Yokoyama, talking about the current situation in which there is demand from a wide range of fields.
SOURE: Yahoo
