Huawei has released the Kirin 9050 series of chips. The most eye-catching one is that a large MoE model with 30B parameters can be run locally on the mobile phone without the need for an Internet connection. Huawei says this is an industry first.
This is achieved by a new generation of DaVinci architecture NPU. The official figures are a 140% performance improvement and a 66% reduction in power consumption.
What is the concept of 30B. At present, most of the mobile phone end-side models are in the 3B to 7B range, which is roughly capable of summarizing, rewriting and simple questions and answers. 30B is close to the size of many cloud mid-range models a year ago. The advantage of the MoE (Mixed Expert) architecture is that only a part of the parameters are activated for each inference. The actual computing power overhead seems much smaller than the 30B figure, but the memory usage cannot be avoided, and the weight of the entire model must be allowed. Huawei did not simultaneously announce the quantization accuracy and memory usage, which determine its actual experience on the 12GB and 16GB models, waiting for real machine evaluation.
In terms of functions, the NPU supports AI screen reading, XMAGE smart photography 2.0, AI eye movement page turning, and screen-off map 2.0.
The 9050 is also Huawei's first chip to use "logic folding" technology. Xu Songyu of Huawei's semiconductor business said that this series has started a new journey for Huawei's chips in the next ten years, and has also set a goal: the P core frequency will exceed 5GHz in 2031.
Privacy and response speed are obvious benefits of running a large model on the device. It can be used in elevators, subways, and airplanes when the network is disconnected, which is a real experience. The difficulty is heating and battery life. If the power consumption reduction of 66% is also true in the scenario of continuous inference, then it will be considered as a real hurdle.
In the next few months, other flagship chips will most likely report even higher numbers.