[ad_1]
By
09.26.2023
0
//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
Ultra-low-power AI Solutions Boost the Capability of Laptop Computers
Laptop computers are an integral part of our daily lives and are used for productivity, collaboration, and entertainment. End-users opt for new laptops with differentiated features that keep pace with smartphones, such as longer battery life, faster system response, intuitive user experience, and enhanced user privacy. According to a recent ARM-commissioned report that surveys desired features for laptops that are available in smartphones, 36% of the respondents wanted longer battery life, 28% wanted faster system response, like instant-on, and 44% were concerned about privacy when working in a public space. To address these feature requirements, laptop manufacturers and technology companies are developing a new generation of ultra-low power machine learning solutions. Himax Technologies is one of the companies that can empower ultra-low power AI to the laptop. Himax’s WiseEye Endpoint AI solution features always-on context awareness that can increase battery life, reduce system latency, enable new user experience, and improve user privacy.
Adopting Endpoint AI Vision Solutions
Power consumption is the main pain point that laptop OEMs or ODMs face when they want to implement always-on applications such as human presence sensing (HPS). Therefore, they adopt mobile and endpoint AI technologies to achieve low power consumption. Current approaches to human presence sensing (HPS) on laptop computers include Time of Flight (ToF) and AI vision solutions. ToF measures the time it takes for light or other electromagnetic waves to travel from a source to an object and back again. While ToF can get 3D image, it consumes more power and requires an extra component.
Vision-based approaches, on the other hand, enable the detection of objects and events using image sensors. The main camera can be re-used as the image sensor in this approach, reducing the system cost. Combining computer vision AI models, vision-based approaches can further analyze the type of moving objects in order to prevent false trigger events and reduce the power consumption due to the false trigger that turns on the SoC. This type of approach, however, requires more computation power because of intensive CNN (Convolutional Neural Network) operations. This inspires IC companies’ endeavors to design AI accelerator-embedded chips.
Although AI vision solutions seem to be a promising approach that allows HPS to be more accurate and versatile, they often consume more power as well. In order to achieve always-on (AoN) sensing capability, AI chips require a unique design to save power. This is how Himax’s WiseEye2 (WE2) becomes a distinctive solution for laptop computers. Facilitated by Arm Cortex M55 CPU and Ethos U55 NPU, the new generation of Himax low-power endpoint AI chip, compared to its previous generation chip, the inference speed is increased by 32 times, and the energy efficiency is increased by 50 times. It allows for the execution of more complex models and obtains more accurate inference results while consuming less power.
Himax WE2 Features
WE2 AI solution is composed of two components: an ultra-low power CMOS image sensor and an AI microcontroller HX6538. HX6538 is an extreme-low power, high-performance microcontroller designed for battery-powered endpoint AI applications. Combining Himax’s ultra-low power CMOS image sensor and HX6538, the WE2 solution allows the device to be always-on and consumes single-digit milliwatt battery power.
WE2 implements multiple strategies and features to support always-on AI solutions, including:
- Efficient neural network processing units
- Muti-layer power management
- Model quantization and pruning
- Security and privacy
- Beyond vision
Efficient Neural Network Processing Units
WE2 is embedded with dual ARM Cortex-M55 CPU cores and an ARM Ethos-U55 microNPU core to accelerate a variety of neural network operations, including convolution and matrix multiplications. Additionally, it supports weight compression so that the model size can be significantly reduced, collectively allowing WE2 to perform complex CNN inference with limited performance degradation.
Multi-Layer Power Management Architecture
WE2 is designed with a multi-layer power management architecture controlled by a hardware state machine. This design reduces the total power consumption. When the first layer, the motion detection block, senses a movement, it will trigger the second layer, Cortex-M55 and Ethos-U55 cores, to run AI models. Given the minimal power consumption of the first layer, WE2 can sustain an always-on state in an extremely low-power mode.
Model Quantization and Pruning
Computer vision tasks often rely on large deep convolution models to achieve high performance. For instance, object detection using ResNet152 demands over 11 gigaflops of computation. This approach, however, is not feasible for edge AI devices due to constrained memory size and the need for short inference time. By implementing model quantization and pruning, the model’s size is significantly reduced while retaining an impressive performance, ensuring a well-balanced solution for edge AI scenarios.
Security and Privacy
Another key value that WE2 brings is its privacy and security features. WE2 not only enables laptops to have AI privacy protection functions, including peak detection and walk away lock, but also enhances the security from the hardware level for both the users and laptop vendors. Most of the time, WE2 performs CNN models that take the image as the input and only transmit the metadata result to the motherboard. This prevents the system from acquiring the user’s image data, greatly improving user privacy. Furthermore, WE2 has a hardware input pin control VMUTE, which allows the user to directly switch off the image output. In addition, WE2 implements Physical Unclonable Function (PUF) and TrustZone security, preventing the device and model weights from being cloned by others.
Beyond Vision
WE2 not only has camera interfaces, such as high throughput MIPI interface and low-power SDI interface, but it also has audio interfaces, PDM and I2S, for audio applications. Similar to the motion detection block for vision, there is a Voice Activity Detection (VAD) block embedded in WE2 to reduce power consumption and enable always-listening use cases. In the future, WE2 can also run vision audio fusion models to achieve better detection capabilities. Himax continues to work on strengthening its AI technology suite, where WE2 can also run vision audio fusion models in the near future to achieve better detection capabilities, broadening a broad array of application possibilities.
[ad_2]
Source link