Charging accessory manufacturer Anker announced the launch of its self-developed chip Thus, and plans to fully introduce local artificial intelligence capabilities into audio equipment, mobile accessories and IoT products. According to the official introduction, Thus is the world's first "compute-in-memory" AI chip for audio scenarios. It is smaller and consumes less energy than traditional chips, and is more suitable for small devices with limited size and limited power supply capabilities.
Steven Yang, co-founder and CEO of Anker, said that most existing AI chips adopt an architecture of “model storage on one side and calculation on the other side.” Each time the device performs inference, a large number of parameters need to be transferred between storage and computing units at an extremely high frequency, which is both time-consuming and energy-consuming. Thus, computing power is deployed directly at the location of the model, "let the calculation be completed where the model is", thus avoiding repeated movement of model parameters inside the chip and improving energy efficiency and response speed.
The first Thus chip will be used in the upcoming flagship true wireless earphones of Anker’s audio brand Soundcore. Anker said that the reason for starting with headphones is that headphones are one of the most challenging product forms for integrating AI chips: the internal space is extremely limited, the battery capacity is limited, and the chip needs to work almost continuously during the period of wear. In the past, it could only run small neural networks with parameters in the order of hundreds of thousands. Taking advantage of the energy efficiency advantages brought by the integrated storage and computing architecture, Thus chips can carry millions of parameters, significantly improve local computing capabilities under similar volume and power consumption conditions, and better cope with tasks such as complex environmental noise.
In terms of call noise reduction, traditional solutions rely on small local neural networks. In strong noise environments, it is often difficult to accurately separate human voices. It is easy for a large amount of environmental noise to leak, or the human voice is severely compressed and distorted, affecting the sense of hearing and call clarity. Anker said that based on the larger-scale neural network available in the Thus chip, coupled with 8 MEMS (micro-electromechanical systems) microphones and 2 bone conduction sensors equipped on the headset to capture the wearer's voice more concentratedly, the new headset, which has not yet been officially released, will achieve cleaner call sound quality in various environments.
However, Thus's actual performance remains to be tested by the market. This integrated storage and calculation AI chip will compete head-on with high-end true wireless headphones including Apple AirPods Pro 3 and Sony WF-1000XM6 in real use scenarios in the future.