Canaan Technology's second-generation AI chip K510 has a computing power of 3TOPS and a power consumption of only 3W

On December 11, Zhang Nangeng, Chairman and CEO of Canaan Technology, announced that the company's second-generation AI chip, K510, has officially been put into production. After two years of research and development, the chip is positioned in the mid-to-high-end market and is an edge AI chip that is completely oriented towards application scenarios. The production of chips means that K510 has entered the trial production stage.

Designed for AP scenarios with a peak computing power of 3TOPS

Kanzhi K510 was developed for AP (application processor) scenarios from the beginning. Its peak computing power can reach 3TOPS, which is three times higher than the first-generation AI chip. Its power consumption is about 3W, continuing the low-power characteristics of the first generation of chips and can meet the power consumption limitations of most devices.

Thanks to the improvement in computing power, Kanzhi K510 will be used in more scenarios such as smart home, smart retail and ADAS assisted driving.

Equipped with a general neural network engine

As the second-generation AI chip, K510 has made significant optimizations in its architectural design. The most obvious adjustment is the use of the GNNE engine to replace the previously developed visual neural network accelerator KPU.

Zhang Nangeng, Chairman and CEO of Canaan Inc., previously stated that when defining the K210, the team did not have a deep understanding of the AI ​​model. MobileNetV2 and YOLOv3 algorithms were not available until the chip was put into production, so the requirement was very simple, which was 1TOPS of AI computing power.

GNNE is a general neural network engine independently developed by Canaan Technology. Compared with KPU, this engine supports larger-scale vector operations and super-resolution. This means that the utilization rate of computing resources of Canaan K510 is greatly improved, achieving higher computing power with the same computing resources.

At the same time, the chip also supports richer peripherals, such as high-definition multi-channel camera interfaces, which can be used for task scenarios such as face recognition and traffic flow analysis in smart stores.

The versatility of the K510 is also reflected in its support for modern machine learning frameworks. The chip's technical specifications will be fully compatible with TensorFlow and support all operators in the TensorFlow file format. In principle, models using TensorFlow do not need to be recompiled or trained, which will provide more convenience for developers.

Zhang Nangeng said that the production of K510 is a milestone for the company. The R&D team has gone through many challenges in the past two years. The production of K510 is also a testimony to a more tenacious and powerful R&D team.

Public data shows that Canaan Technology has maintained a 100% success rate in the past seven chip castings and mass productions.

-End-