Wearables Demand Dedicated NPUs for AI/ML Operations

Hardware and Computing Trends in the Wearables Market

Since the invention of the microprocessor in the 1970s, the overarching trend in computing has been to lower the barriers between the user and the computer. From original computing systems that occupied entire rooms to smartphones that allow you to tuck computing into your pants pocket, the computing industry is in a constant battle to bring humans and devices physically closer together.

Today, wearable devices allow you to wear your computing resources. Between smartwatches and the recent emergence of augmented reality and virtual reality (AR/VR) headsets, research projects that the global wearables market will reach $186.14 billion by 2030, up from $71.91 billion in 2023. 

However, before this future can be realized, engineers must conquer several notable challenges. 

Machine Learning at the Edge

Like most other fields of technology, one of the biggest trends in wearables is the proliferation of machine learning (ML). However, unlike many applications, like ChatGPT, which rely on cloud computing resources, wearable products face the unique challenge of performing machine learning computation at the edge.

In technical terms, edge computing enables the execution of ML models directly on wearable devices, using onboard processing capabilities to analyze and infer from data locally, without requiring constant communication with cloud servers. 

From an operational perspective, embedding ML at the edge requires integrating compact, efficient ML models into the device firmware. Techniques such as model pruning, quantization, and knowledge distillation are employed to compress and optimize ML models, ensuring they fit the constrained environments of wearable devices. In traditional cloud-based environments, the ML model execution is done on specialized hardware accelerators like GPUs (Graphics Processing Units) or TPUs (Tensor Processing Units), however, such solutions are usually too large, power hungry and expensive for use in a true edge device. In such systems, a microcontroller with an integrated NPU is a much better option, as they are optimized for on-device ML workloads, enabling faster processing while minimizing power consumption.

A man looking through a virtual reality device attached to his head, while interacting with the images he sees.

AR/VR applications rely on real-time, edge computing. (Source: Adobe Stock)

The impetus for machine learning to operate at the edge is primarily dictated by the requirements for real-time processing, privacy preservation, and energy conservation. 

Real-time processing is essential for applications such as health monitoring and augmented reality, with immediate feedback or action required. For example, in an AR setting, real-time object recognition and spatial mapping are essential for overlaying digital information onto the real world seamlessly. Whereas cloud solutions have round-trip latency delays, an edge-based ML model could instantly analyze the camera feed to identify objects and spaces.

Edge computing further enhances privacy and data security, as many of these devices deal with sensitive data, like biometrics or personal health information. In a cloud computing scheme, this sensitive data must be physically sent to the cloud location, opening up avenues for interception and vulnerabilities. With edge computing, this data is processed on-device, significantly reducing the risk of interception during transmission to the cloud. 

Bluetooth Low Energy (BLE)

A defining characteristic of most wearable devices is that they’re connected devices: they have a means of wireless communication with a larger network. For example, a health monitoring smartwatch shares biometric data with a centralized dashboard for analysis. Or, imagine an AR/VR headset that needs to communicate with a mobile app or synchronized hand controllers.

In this context, Bluetooth Low Energy (BLE) has become a prominent wireless protocol in wearables. The technology’s ability to facilitate reliable, continuous, and energy-efficient communication between devices is necessary so that wearables maintain longer battery life while providing constant connectivity.

a man controls a robot using a virtual reality device and controllers.

AR/VR devices rely on BLE to connect with peripheral devices like controllers. (Source: Adobe Stock)

The emergence of edge ML in wearables also has a notable impact on the feasibility of BLE in wearable applications. This local processing reduces the need for continuous data transmission and reduces the volume of transmitted data, preserving both bandwidth and battery life. For example, an AR wearable could use on-device ML to recognize user gestures for interface control, and then utilize BLE for low-energy data transfer to execute commands or share insights with linked devices.

Moreover, advancements in BLE, such as support for the Low Complexity Communication Codec (LC3) and features like LE Audio’s Auracast, are revolutionizing audio experiences in wearables. AR/VR devices can now offer personal audio guides or shared listening experiences without compromising on energy efficiency. This is particularly transformative for educational or collaborative AR/VR applications, in which audio clarity and battery longevity are paramount.

Power Efficiency

For a developer, one of the greatest challenges in wearables is balancing the demand for greater performance with the demand for power efficiency.

Greater performance can be thought of as more capable processing, more advanced sensing technologies, and even greater resolution displays. While each of these is necessary for improving user experience and overall device functionality, they also come with a trade-off: power consumption. All other things being equal, achieving more performant computing, greater sensing capabilities, or better display technology often requires more electrical power, which poses challenges for wearables.  

The first challenge is with battery operation. Most wearables are standalone battery-powered devices, meaning that their runtime is only as long as the battery life. As devices become more performant and more power-hungry, the battery life significantly diminishes. What good is an extremely performant AR/VR headset if its battery lifetime is too short to be useful? Unfortunately, the answer isn’t as simple as adding a larger battery, as weight, cost, and space constraints don’t always allow such a solution.

Another corollary of higher power consumption is increased thermal challenges. As devices burn more electrical power, they become notably hotter. For wearable devices like AR/VR glasses which get worn on the head, or smartwatches which are on the wrist, high temperatures are uncomfortable for users and can even pose safety risks. 

All things considered, navigating these design trade-offs hinges on developing power-efficient hardware that can achieve greater performance without sacrificing on power consumption. 

On the hardware side, power efficiency in wearables is often achieved through advanced microprocessors, which optimize power consumption by operating at low voltages while maintaining high performance levels. These processors often feature dynamic frequency and voltage scaling, allowing them to adjust their power usage based on the workload, thereby conserving energy during less-intensive tasks. Further, the integration of specialized processors for neural networks improves the power efficiency of machine learning tasks. 

Lowering Barriers

The widespread adoption of wearables is the next major step in the world of computing, yet major challenges remain. The emergence of on-edge machine learning, the proliferation of Bluetooth Low Energy, and the push for more energy-efficient hardware resources are keeping the industry hard at work. For AR/VRs and other wearables, advancements in microcontrollers and processors are paving the way for better products, with improved features such as machine learning at the edge, BLE, and power efficiency. 


This field is for validation purposes and should be left unchanged.