While the medical segment is expected to have the fastest growth over the forecast period, the automotive and the transportation segment will register the most lucrative growth, AMR notes.

Despite these bullish forecasts, however, the AI processors segment continues to face challenges, especially when it comes to performance and power. And for endpoint AI applications, the main issues include providing powerful AI computing capabilities under low power constraints.

For instance, maintaining extremely low standby power while having fast wake-up and peak compute; integrating high performance hardware within limited size and cost budgets; and optimizing software and models to run on resource constrained devices. This requires holistic innovation across algorithms, chip design, software-hardware synergy, and more.

“Key challenges faced by our customers in AI-based sensing include power consumption, latency, model efficiency, and optimizing hardware and software,” says Alex Chang, Product Manager at Himax Technologies Inc. “Devices need to balance high accuracy with ultra-low power for continuous sensing.”

Himax Technologies is a fabless semiconductor company specializing in display drivers, timing controllers, image processors, and AI processors. According to Chang, the company is one of the leaders in visual AI solutions.

Helping customers through efficiency, device architecture, and optimization

“We help our customers address these issues through the revolutionary efficiency, specialized heterogeneous architecture, and software optimization of our endpoint AI solutions,” says Chang. “One example is our WiseEye 2, an ultra-low power AI processor optimized for on-device computer vision. Key advantages are the 60mW facial mesh processing, dedicated microNPU core, advanced power management, abundant on-chip memory, and open software platform.”

WiseEye 2 is suitable for applications requiring real-time responsiveness and decision making—such as industrial quality inspection, predictive maintenance, autonomous driving. It also fits applications with strict demands on power and size, like wearables. Additionally, its open software platform facilitates developers to create new applications.

As its name implies, WiseEye 2 is the second generation of Himax Technologies’ AI accelerator-embedded ASIC platform for ultra-low power applications. Its biggest innovation over its predecessor is the dedicated 400MHz Ethos-U55 microNPU optimized for machine learning workloads, which can accelerate operations and greatly improve efficiency.

Additionally, the heterogeneous multicore design with the energy-efficient 150MHz Arm Cortex-M55 CPU, high-speed 400MHz Cortex-M55 CPU, and microNPU can match real-time workloads. The Advanced Power Optimization Techniques with Layer-By-Layer wake-up allows minimal power consumption by intelligently activating components. Its abundant on-chip memory holds entire neural networks, avoiding repeated accesses. Together, these allow the highly complex 468-point facial mesh analysis using only 60mW of power—a 10-100X efficiency improvement over previous solutions.

According to Chang, WiseEye 2 employs advanced techniques such as voltage/frequency scaling, sleep modes, power islands, and Layer-By-Layer wake-up to minimize power usage. On the other hand, the dedicated microNPU, heterogeneous multicore CPU design, and parallel execution, meanwhile, help optimize performance.

“The extensive specialized hardware and software co-design enables it to deliver the highly complex 468-point facial mesh analysis in just 60mW—powerful AI performance at an extremely low power budget that other solutions cannot achieve,” says Chang.

What’s next for Himax

WiseEye 2 embodies the direction of achieving efficient endpoint AI via synergistic algorithm, hardware, and software design, according to Chang.

“The intelligence of edge devices will continue improving, with more cloud capabilities being migrated on-device. Areas of focus include improving compute efficiency, integration of memory and sensors, and end-to-end software-hardware optimization,” he says.

WiseEye 2 represents the leading-edge of endpoint AI processor technology, pushing the envelope in power efficiency, performance, and intelligence. Chang notes that its biggest advantage is the extensive software-hardware co-design and optimization, which gives it superior price-performance compared to other solutions. Additionally, the complete solutions and development tools are also key competitive strengths.

“Our roadmap includes further improvements to embedded AI efficiency, always-on sensing capabilities, and tinyML model optimization. We aim to push the boundaries of what is possible on endpoint devices,” says Chang. “We are planning two different generations: one for the notebook business, which is the current ongoing project, and a next-generation device for IoT applications.”

Bullish outlook

“I see strong demand for on-device AI continuing, driven by edge computing and IoT,” says Chang. “Specialized AI processors like WiseEye 2 that push efficiency boundaries will be key to enabling many AI applications at the endpoint.”

For example, in industrial manufacturing, edge computing enables process optimization, quality control, and predictive maintenance by continuously analyzing sensor data on-premise rather than sending large volumes of IoT data to the cloud—this reduces costs and latency.

“Autonomous vehicles can make split-second maneuver decisions using real-time sensor analysis rather than waiting for the cloud to respond–instant response from edge computing can be critical,” says Chang. He also adds that in retail environments, edge nodes allow products to be instantly recognized using computer vision, enabling new frictionless pickup and checkout flows for customers.”

Meanwhile, from a medical/healthcare standpoint, patient health can be continuously monitored by hospital edge nodes for early intervention and therapy adjustments based on real-time data analytics. This can improve patient outcomes. Finally, smart city infrastructure can manage traffic flows, public safety, environmental quality in real-time by using distributed edge intelligence to analyze data from city-wide sensors and feeds.

“I’m excited about the future possibilities unlocked by bringing ever more sophisticated AI capabilities to the edge through purpose-built solutions like WiseEye 2. Embedded intelligence will redefine what’s possible across many industries,” Chang concludes.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved