As AI sinks towards the terminal, end-to-end AI makes intelligence within reach. How to view the development trend of end-to-end AI, and why does end-to-end AI play a key role in the vision of "an AI future for everyone"? Xinhua News Agency interviewed Wan Weixing, the head of Qualcomm's AI product technology in China. The application of generative AI is increasingly being implemented on various terminal sides, and the trend of AI localization has been recognized by chip manufacturers, OEM manufacturers, ODM manufacturers, and upper level application manufacturers. Wanxing believes that compared to the cloud, the advantage of end-to-end AI is very obvious. In addition to cost advantages, it also has unique advantages in privacy protection, more personalized user experience, and anytime access to AI. The deployment of generative AI on the end side has brought tremendous changes to work and life, which is why only by deploying generative AI on the end side can we help standardize and expand the use of generative AI. AI is sinking towards the edge, posing significant challenges to terminal side chips and software. Seemingly simple use cases such as AI image recognition, voice wake-up, and computational photography actually have strong demands on the computing power and DDR bandwidth of the chip, making it difficult to design a hardware IP (processor) that can meet all these applications. "From our perspective, for sudden real-time tasks, we will use universal hardware acceleration units, including CPUs and GPUs, to accelerate these use cases. For sustained and high computing applications, such as photography, video recording, and generative AI, we will recommend running on our high computing power, low-power NPU processors. If you are familiar with the Snapdragon platform, we also have an ultra-low power sensor hub, which is particularly suitable for some use case scenarios that require always on." Wanxing summarized that Gaotong will take advantage of its advantages on the chip, utilize the technical expertise of the processor, and design a complete heterogeneous computing system. This heterogeneous computing system The computing system will include all processors for different use cases, Using different processor components for acceleration to enhance user experience. Qualcomm integrated the concept of AI throughout the entire SoC from the very beginning of its design. At the 2023 Snapdragon Summit, Qualcomm demonstrated a use case of utilizing heterogeneous computing for acceleration, which is a very typical scenario that can fully tap into the computing power of heterogeneous computing on the Qualcomm Snapdragon platform. The use case we presented at the time was a scenario of conducting interviews or conversations with virtual humans, which involved many technologies. Firstly, we needed a speech processing module to convert speech signals into text signals, which could be run on a CPU or an ultra-low power sensor hub. Next, we needed to use a large model to generate dialogue output text, which could be run on an NPU. Then, to complete the process of voice dialogue, we also needed a module to convert text back into speech, which could be accelerated using a CPU or GPU. Finally, in order for both parties to see each other, we needed a rendering engine to display their images in the same scene. Rendered in, this can be utilized