UMEC(HK)’s Reconfigurable Chiplet System Is in Place for High Performance, Low Power Edge AI Applications
As one of the breakthroughs in the semiconductor industry, chiplet-based design technology has changed the industry ecosystem by integrating multiple chiplets with different functions in the same package. UMEC(HK) recently released a chiplet system which is widely applicable to different edge AI applications. Compared to traditional systems with discrete components, it is much smaller, energy-efficient and supports shorter time to market.
The complete all-in-one system is formed by connecting chiplets with various functions in tiny package (9.5mm x 6.8mm). It provides field-programming features that allow rapid reconfigurations at multiple levels, ranging from chip-to-chip level to AI core level. By supporting diversified convolutional neural network (CNN) models, developers can apply audio and visual algorithms easily. The system is highly compatible with different types of smart sensors, opening up more possibilities for the AIoT applications.
A customised CNN model is developed to pair up with the chiplet system, further enhancing the performance with ultra-low power consumption. Applying suitable CNN model to chiplet systems can be challenging as AI algorithms based on CNN normally involve huge computation resources and massive memory storage, but limited hardware resources such as memory capacity, logic gate number, etc. are available in these systems. Under the restrictions, our CNN model utilises extremely low-bit quantisation method to achieve more than 16-time weight reduction. More than that, our patented compression-retain technique can further reduce 20% memory usage without accuracy loss. Using the algorithm-hardware co-design method, we successfully developed a micro-watt level solution which is suitable for different portable AI computing devices.
The chiplet system provides a perfect foundation for a variety of AIoT applications, for example, voice-controlled office automation solution supported by keyword spotting and voice activity detection, intelligent health monitoring solution with human detection, body temperature sensing and thermal imaging, or any solutions that you can imagine. It has unlocked countless possibilities for edge AI applications, making the impossible of the past possible today.