On the 25th, it was learned from Nanjing University that Professor Miao Feng's team at the university's Brain Intelligence Technology Research Center proposed and validated a new "End Cloud Fusion" intelligent computing paradigm. This paradigm integrates wireless communication into neural networks, significantly reducing the power consumption of wireless communication modules while maintaining high-precision inference capabilities. It provides a new theoretical perspective and engineering path for the intelligent interconnection of ultra large scale end side devices. The relevant research results were published on February 25th in the international academic journal Nature Electronics. For a long time, a large number of end side devices such as embodied robots have had limited computing power and require collaborative intelligent computing with cloud devices. However, transferring massive amounts of data to the cloud for processing consumes too much energy and incurs high costs, making it difficult for edge intelligence to be widely implemented. In the era of the Internet of Things, building a cloud based wireless collaborative network urgently requires the development of a new collaborative paradigm to break through the communication energy bottleneck caused by traditional lossless data transmission. ”Miao Feng stated that this paradigm requires lower transmission power and higher computational energy efficiency to support intelligent interconnection of ultra large scale terminals. In response to this demand, the research team used a self-developed simulated in memory computing chip to construct a neural network inference system and a wireless communication system, which were respectively used for the inference calculation of the end side neural network model and the wireless transmission of the calculation results. Simulated in memory computing technology has the characteristics of highly parallel computing, which can achieve high complexity matrix operations in neural network inference and communication signal processing. ”The co corresponding author of the paper and assistant professor at Nanjing University, Wang Cong, explained. At the level of training methods, the research team proposed a "algorithm hardware" collaborative optimization approach for communication aware training, which significantly reduces the energy consumption and hardware costs of wireless communication and improves the robustness of the system. By incorporating 'wireless communication' into the optimization training of neural networks, the system will actively learn 'how much energy should be spent to transmit data' while ensuring the accuracy of inference tasks. This allows the end cloud fusion computing system to complete high-precision inference tasks at a lower energy cost in various wireless communication environments and modulation methods. ”Introduction by Liang Shijun, co corresponding author of the paper and professor at Nanjing University. This is particularly important for intelligent terminal devices in complex and ever-changing wireless environments in mobile scenarios. ”Miao Feng stated that this research breaks through the design formula of traditional end cloud collaborative computing systems and proposes a new paradigm of intelligent computing that is "task centric and end-to-end collaborative optimization", providing a new path for efficient intelligent computing of ultra large scale terminal devices. (New Society)