Industrial automation, robotics, smart cities, and home automation are just a few of the application areas driving the demand for increased computing performance and capabilities. In the past, sensors in these types of systems were much simpler and not interconnected; however, artificial intelligence and machine learning (ML) are now able to enhance local intelligence, allowing for local device decision-making that was impossible with simple control algorithms in the past.
The Evolution of General-Purpose Processors in the Age of Artificial Intelligence
Years ago, developers focused on logic and control algorithms as the core of software development; however, the emergence of digital signal processing (DSP) algorithms has made many enhanced speech, vision, and audio applications possible.
This shift in application development has ushered in a new era, influencing the design of computing architectures. We have now reached a stage where inference is a major focus of algorithm development. This brings new or higher requirements for computational performance, energy efficiency, latency, real-time processing, and scalability.
We see that not only are new processor accelerators needed, but also improvements in general-purpose processing, providing developers with the necessary balance and enabling applications such as feature inspection or real-time video detection of people.
A few years ago, when developers created noise cancellation applications, they relied solely on frequency-based filters. Today, however, developers can improve application performance and functionality by blending filtering with ML/AI models and inference. This has driven demand for processors and tools that make these tasks more efficient and as seamlessly accessible as possible to users.
Smart tags in edge and endpoint devices
This evolution is driven by ML, but it is not without technical challenges. Years of startups and shutdowns, as well as attempts to create a "one-size-fits-all" approach, have prompted the industry to change its methodology in order to unlock opportunities for massive scaling.
Developers are now leveraging secure and performance-enhancing technologies to enable small, low-power embedded systems for previously unimaginable sound, vision, and vibration applications that are changing the world. Various versions of language and transformer models will soon find their place in IoT edge devices with new computing capabilities. This will open up new possibilities that developers have never dreamed of.
To equip developers with the hardware needed for this development transformation, we introduced ARM vector processing technology into the ARM8.1M architecture several years ago. As a small, low-power embedded device, ARM offers significant performance improvements in ML and digital signal processing applications. It also provides Single Instruction Multiple Data (SIMD) functionality, delivering a new level of performance for ARM Cortex-M devices and supporting applications such as predictive maintenance and environmental monitoring.
ARM enhances the performance of DSP and ML, accelerating signal conditioning (such as filtering, noise reduction, and echo cancellation) and feature extraction (audio or pixel data), which can then be fed into the classification part via a neural network processor.
Achieving intelligent edge capabilities
We have seen partners choose ARM technology in their latest products, enabling developers to leverage the ML capabilities of the most constrained devices at the edge of the network. The ARM Cortex-55 was released in February 2020, while Alif Semiconductor released its first Cortex-55 silicon in September 2021. It uses ARM's Cortex-M-55 in its overall product lineup and product family. HMAX has also adopted the ARM-powered Cortex-M-55 as part of its next-generation W2AI processor, which targets computer vision in battery-powered IoT devices.
The second CPU to use is the ARM Cortex-M85, released in April 2022. Renas showcased the M85 at Embedded World 2022 and 2023. Plummer, a company that develops complete software solutions for camera-based person detection, utilized Renas' RAMU technology to significantly accelerate its inference engine during the demonstration. The company believes that the performance improvements will enable its customers to use larger, more accurate Plummer person detection AI, add additional product functionality, and extend battery life.
As hardware continues to evolve, developers face increasingly complex software requirements, necessitating new development processes to create optimized ML models and efficient device drivers. Crucially, the software development platforms and tools provided to the ecosystem must evolve alongside the hardware.
There are a wide variety of ARM and third-party tools available to support end-users in creating AI algorithms. Once data scientists have created the model offline, tools are available to optimize it, whether on an NPPS-based processor or on a Cortex-based processor using ARM instructions.
QEexo was the first company to automate end-to-end machine learning for edge devices. Its automation platform provides an intuitive user interface that allows users to collect, clean, and visualize sensor data, and automatically build machine learning models using various algorithms. Traditional embedded tools, such as the Key Microcontroller Development Kit (Key MDK), complement the MLPOP tool and help establish a development flow for validating complex software workloads. Thus, embedded, IoT, and AI applications converge in a single development flow, a concept known to many software developers.
The potential of the edge is being unlocked. There is a growing demand for microcontroller performance, particularly in areas such as voice-activated door locks, personnel detection and recognition, connecting motor control with predictive maintenance, and countless other high-end artificial intelligence and ML applications.
With the right technology, developers can reimagine edge and endpoint devices, striking the right balance between cost, performance, energy efficiency, and privacy—key factors for these constrained systems.
