Arm has recently introduced its latest and most potent microcontroller-grade neural processor, the Ethos-U85. This cutting-edge processor is capable of reaching up to 4Top/s in its maximum configuration, making it a significant advancement in the field. What sets the Ethos-U85 apart is its impressive 20% increase in power efficiency compared to its predecessors, showcasing Arm's commitment to pushing the boundaries of neural processing technology.
Named Ethos-U85, this innovative processor is configurable with a range of between 128 and 2,048 MACs. The primary objective behind its development is to provide sufficient performance to enable the local execution of complex AI algorithms, including 'transformers' and convolutional neural networks. According to Arm, transformer networks are poised to revolutionize various applications, particularly in the realms of vision and generative AI, enabling tasks such as video comprehension, image completion, and multi-camera data analysis for image classification and object detection.
Arm's toolchain for the Ethos-U85 offers support for popular AI frameworks like TensorFlow Lite and PyTorch, enhancing its versatility and ease of integration with existing AI ecosystems. The 4Top/s performance metric, achieved at 1GHz with int8 data, positions the Ethos-U85 as a significant leap forward from its predecessor, the Ethos-U65, which delivered 1Top/s. In comparison, the Ethos-N78 could reach up to 10Top/s, highlighting the remarkable progress in neural processing capabilities.
Given its impressive performance metrics, the Ethos-U85 is expected to complement higher-end Cortex-M CPUs and lower-end Armv9 Cortex-A CPUs, offering a comprehensive solution for a wide range of AI applications. In conjunction with the Ethos-U85 launch, Arm also unveiled a reference design that pairs the neural processor with the high-performance Cortex-M core, the M85. This design, known as Corstone-320, features additional components such as the Mali-C55 image processor and a new DMA controller, DMA-320, aimed at delivering exceptional performance for voice, audio, and vision applications.
The Corstone-320 reference design is designed to support real-time image classification, object recognition, and natural language translation for voice assistants on smart speakers. Arm's commitment to providing a holistic solution is evident in the inclusion of software, machine learning models, and tools, as well as virtual hardware for code development. With the Ethos-U85 expected to be integrated into ICs next year, the possibilities for its applications are vast, ranging from small library language models for local language identification to industrial vision-based defect detection using transformer networks.