New Attempt in Machine Learning: Making Neural Processing Units Mainstream

2019-10-24

ARM is an intellectual property (IP) supplier that differentiates itself from typical semiconductor companies by not manufacturing or selling chips to end users. Instead, it transfers design schemes to partners who produce unique chips. Through this mutually beneficial partnership, ARM has quickly become the creator of global RISC microprocessor standards. Recently, the company has set a grand goal of introducing machine learning into mainstream and even low-end devices through its new neural processing units (NPUs). ARM will launch the Ethos-N57 and Ethos-N37 NPUs and license them to chip manufacturers for integration into their products. The aim is to expand the range of ARM machine learning processors and enable the application of artificial intelligence in mainstream devices. Earlier, the company introduced the Mali-D37 display processing unit (DPU), which offers a rich set of display features and enables full HD and 2K resolutions in minimal areas. Paul Williamson, Vice President and General Manager of ARM's Client Line of Business, stated in a declaration that in the near future, AI will no longer be exclusive to high-end devices. Nowadays, artificial intelligence is ubiquitous, from gaming devices to digital televisions (DTV), its presence can be seen everywhere. Therefore, there is a need to increase computing power in terminals to meet customer expectations. For example, smart experiences in DTV should include not only basic intelligent assistant voice commands but also real-time translation of foreign language programs and convenient facial recognition systems. ARM Ethos is a product suite designed to address complex challenges in artificial intelligence and multilingual computing, allowing for more personalized and immersive experiences in everyday devices. As consumer devices become smarter, they require additional AI performance and efficiency through dedicated ML processors. The new Ethos NPU is optimized for cost and battery life-sensitive designs, providing high-quality artificial intelligence experiences on everyday devices. The suite of IP includes three components: - Ethos-N57 and Ethos-N37 NPUs: Supporting artificial intelligence applications, balancing ML performance with cost, area, bandwidth, and battery life limitations, and enabling true heterogeneous computing. - GPU: The first mainstream GPU based on the Valhall architecture, providing users with an immersive experience through performance improvements. - Mali-D37 DPU: Offering rich display features in minimal areas, making it the perfect DPU for entry-level devices and small displays. Among them, Ethos-N57 and Ethos-N37 are the foundation, designed with considerations for int8 and int16 data types and utilizing advanced data management techniques to minimize data movement and related power consumption. Innovative implementations such as Winograd have improved their performance by over 200% compared to many other NPUs. Williamson openly states that the new IP represents ARM's expanding capabilities, bringing high-quality experiences to everyday and highly efficient consumer devices.

RECENT AI NEWS

RECENT AI TOOLS