AI chips, also known as AI accelerators or compute cards, are modules dedicated to handling a large number of computational tasks in AI applications (other non-computational tasks are still handled by the CPU), according to Rexroth.
The computing scenarios for AI chips can be divided into cloud AI and endpoint AI. The computational scenarios for deep learning can be divided into three categories, namely, training in data centers, inference in data centers, and inference in embedded devices. The first two can be summarized as applications for the cloud and the latter can be summarized as applications for the endpoint.
Mobile AI chips are able to process various AI operators at a speed of about 30x to 50x, and are able to do better an image detection, image segmentation and image semantic understanding for photo scenes, for example. In addition, the voice can be heard and understood, and the service that the customer really wants is provided based on the understood intent of the customer.
In 2017, Huawei was the first in the industry to release the Kirin 970 processor, Huawei's first artificial intelligence end-side chip, which solves the performance and power consumption problems of running AI models on the end-side from the hardware level, which gives smart terminals powerful computing capabilities.
Taking image recognition speed as an example, Kirin 970 can reach about 2005 images per minute, while it can only process 97 images per minute without NPU. Compared with the previous generation, the graphics processing performance is improved by 20% and energy efficiency is improved by 50%, which can support smooth operation of 3D large games for longer time and support new generation mobile internet experience such as AR/VR.
With the powerful mobile computing power of dual NPUs, Kirin 980 achieves 4500 images per minute, which is 120% faster than the previous generation and much higher than the same period in the industry.
Constructing a platform and playing smart combinations
Rexroth said the future is the integration of four aspects such as cloud, pipeline, AI chips and terminal capabilities. The cloud side and the end side can do better synergy, can exchange capabilities more quickly, and do coordination on capabilities. We have to have such an open platform that can interface all the capabilities within our ecosystem.