Der Blätterkatalog benötigt Javascript.
Bitte aktivieren Sie Javascript in Ihren Browser-Einstellungen.
The Blätterkatalog requires Javascript.
Please activate Javascript in your browser settings.
18 Elektronik International 2020 Microcontroller One size of code fits all The concept of taking a pretrained model and optimizing it for target architectures with different levels of processing performance is gaining traction in the electronics market There is a growing number of models available targeting the Three Vs Video Voice and Vibration as well as image recognition By starting with a pretrained model and passing it through subsequent levels of optimization engineers can arrive at a deployable system much quicker Examples of this crossplatform approach to development are beginning to emerge throughout the industry intended to simplify the developers task of deploying code across potentially multiple architectures including GPUs CPUs AI processors and FPGAs Arm is also a proponent of portable code exemplified by CMSIS its vendoragnostic hardware abstraction layer for microcontrollers based on the Arm Cortex family As outlined earlier this now includes CMSIS-DSP and CMSIS-NN two software libraries that are supporting many of the vendors own code frameworks targeting AI and ML in endpoints such as NXPs eIQ The efforts being made now to port neural networks and ML models to lowpower MCUs in order to empower AI directly in the endpoint illustrate how important and enabling AI will be taking the IoT to an entirely new level and allowing a new tier of end user to access the technology But the efforts arent going to stop with new software frameworks The next steps in AI processing The market for AI processors is estimated by analyst Allied Market Research to exceed $90bn by 2025 which is a good indication of the amount of R&Dcurrently going into developing new processor architectures better designed to run neural networks accelerate AI training and execute ML models Of course as demonstrated any processor can run AI to some extent while those with DSP extensions will have an advantage over those that dont However while DSP instructions are beneficial there are other types of vector processing instructions for compute architectures coming to the market that will provide even greater boosts to executing DSP and AI in endpoint devices This is illustrated by the introduction of Arms Helium technology Helium is part of the Armv8 1-Marchitecture and brings vector processing capabilities to the smallest devices increasing the performance of signal processing functions by five times and delivering a 15x improvement in ML functions Dr Dominic Binks VP of Technology at Audio Analytics has described Helium as changing the game Audio Analytics already uses ML in its products running on Arm Cortex-M MCUs but predicts that optimizing just a few selected routines for Helium would see more than a 50% reduction of execution time With even more effort this execution time will become even faster The key to deploying AI and ML is scalability from the processors resources to the number of heterogenous processors in one device through to the ease of stitching multiple multicore devices together in a single system Only Arm is approaching this holistically to offer a fullscale selection of processing solutions Arm is also actively developing entirely new neural processing architectures and NPUs to address AI and ML requirements the Arm Ethos-N Series currently features the Ethos-N77 Ethos-N57 and Ethos-N37 The N37 is aimed at smaller most costsensitive endpoints but is able to deliver 1TOP s in an area of just 1mm2 Other examples of NPUs designed for low power and low cost are also in the development pipeline This illustrates the commitment Arm is making to developing scalable platforms in addition to the range of Arm CPUs available that will continue to enable ML in the endpoint devices and the many benefits this will have for a modern society FR Thomas Lorenser is director of new products in the Automotive and IoT Line of Business at Arm leading a team responsible for the definition of Arms roadmap for IoT and embedded applications Thomas has 15 years of experience in various Product Management and Product Line Management roles in the semiconductor industry across many market segments including mobile tablets notebooks and wearables Prior to joining Arm Thomas was Product Line Manager at Knowles Electronics and Product Manager at NXP and has extensive global experience from working with OEMs and silicon partners Figure 5 The many layers of a CNN as implemented using CMSIS-NN photo Arm