Thinking deeply about Zou Tianqi every moment, implementing low-power visual gesture recognition using sensory memory simulation chips

    In signal processing such as speech and vision, the traditional signal processing path is from analog signals to analog-to-digital conversion, and then to digital processing. In this way, analog-to-digital conversion has drawbacks such as low computational efficiency and high power consumption. However, analog computing, represented by every moment of contemplation, can preprocess some digital signal processing tasks by performing feature extraction and in memory analog computing in the simulation field, This simplifies the information to be processed. In addition, through preprocessing, ADC does not need to be normally open, which can further reduce power consumption.

    Analog computing can break through the development dilemma of digital chips, so there is great potential for future development. Zou Tianqi stated that at present, Moore‘s Law is approaching its limit and there is a lack of advanced manufacturing processes. At the same time, Feng Neumann‘s traditional storage and computing separation architecture has begun to face storage wall bottlenecks, and the cost of computing power is also high. This has given new development space for simulation computing.

    The thought-provoking simulation calculation pushes memory computing to the extreme, directly performing calculations within the memory array. This can be achieved by using memory elements as tunable resistors, providing input as voltage, and collecting output as current. This means that simulation calculations can be directly used for core neural network matrix operations.

    On the one hand, simulation computing eliminates memory data movement, and on the other hand, it is parallel computing in vector operations, making it easy to achieve high performance and efficiency.

    In memory computing can control the location of data in memory, rather than building a cache hierarchy. In memory computing adds calculations to each memory array to minimize data movement.

    Based on this concept, Everthink has launched a low-power perception chip, offloading gesture interaction computing tasks from the application processor to its dedicated chip, achieving lower power consumption and smaller computing power requirements, and simplifying the addition of functions. Supports coordinate detection of 21 key points in both hands, and supports custom gestures. Its typical power consumption is 200mW, the processing frame rate is greater than 60fps, and the accuracy rate exceeds 98%.