The industrial chain still has shortcomings, but design architecture can become the strong point of domestic AI?

"If we only look at the design architecture, domestic artificial intelligence chips are no worse than foreign ones, but this is not enough." At the "Science Cafe" science salon held by the Institute of Physics of the Chinese Academy of Sciences recently, Wang Yu, a professor at the Department of Electronic Engineering of Tsinghua University, pointed out that this is equivalent to building a building on sand, and the foundation is not solid.

The reporter then interviewed Gong Guoliang, deputy director of the Brain-Inspired Computing Research Center of the Institute of Semiconductors, Chinese Academy of Sciences, who expressed a similar view.

The design architecture is remarkable

Wang Yu participated in the establishment of the artificial intelligence chip company SenseTime in 2016. Now more companies engaged in artificial intelligence chips have emerged in China. He feels that their performance is "no worse than that of foreign companies, and some are even better."

The "good" mentioned here mainly covers the design architecture level. There are certain reasons why the design architecture has become the strong point of domestic artificial intelligence chips.

"Traditional CPU and GPU chips use a von Neumann computing architecture and computing model based on instruction stream to operate, while artificial intelligence chips mostly use brain-like or brain-imitation architecture, which can break through the computing bottleneck of the memory wall." Gong Guoliang introduced that in scenarios such as speech recognition and image recognition that use deep neural network algorithms, artificial intelligence chips have significant advantages over traditional chips.

Due to the above-mentioned differences with traditional chips, Gong Guoliang told Science and Technology Daily that from the perspective of design architecture alone, domestic artificial intelligence chips can quickly keep up with related technologies and are almost at the same level as developed countries.

Recent news provides us with a glimpse into the rapid progress of domestic artificial intelligence chips.

On June 20, Cambricon launched its second-generation cloud-based artificial intelligence chip "Siyuan 270"; on June 21, Huawei released the artificial intelligence mobile phone chip "Kirin 810"; on July 3, Baidu released the far-field voice interaction chip "Honghu".

Gong Guoliang introduced that most of the AI ​​chips developed by Huawei, Cambrian and other companies are general-purpose chips that can handle a variety of AI application scenarios. In addition, many universities and research institutes are also developing AI chips, most of which are dedicated chips with strong performance in special application scenarios.

The industrial chain still has shortcomings

Although the development of domestic artificial intelligence chips is in its infancy and has made remarkable achievements in design architecture, experts point out that China's overall foundation in the chip field is weak and there are still many constraints and shortcomings.

Wang Yu introduced that the latest equipment and technology for chip manufacturing in mainland China lags behind the international advanced level by one to two generations. Therefore, some artificial intelligence chips need to be sent abroad for manufacturing and packaging, and the completeness of the industrial chain is poor.

Gong Guoliang confirmed the above situation in an interview. He told Science and Technology Daily that the artificial intelligence chips used in mobile phones are a typical example. These chips often need to use the latest technology to reduce power consumption, improve integration and computing performance. They are high-end chips with a very large demand, but mainland China does not yet have the conditions for manufacturing and packaging.

The weaknesses of the development of domestic artificial intelligence chips include high-speed interfaces and dedicated integrated circuit IP cores. Gong Guoliang introduced that, taking the latter as an example, they have relatively high computing performance, complex design, and are closely related to the manufacturing process. It usually takes years of accumulation to realize such a design module.

"High-speed interfaces and dedicated integrated circuit IP cores are regarded as core technologies by the industry. Using them often requires authorization and licenses from foreign companies, and companies that do not possess these technologies will find it difficult to obtain licenses in the short term," said Gong Guoliang.

To summarize in Wang Yu's words, domestic artificial intelligence chips generally perform well in the aspects that require intelligence and wisdom, but not well enough in the aspects that require accumulation and sedimentation.

"Overall, research in related fields in China started late and does require accumulation and sedimentation." Gong Guoliang believes that in the long run, China's artificial intelligence chip field should seize the opportunity to make up for weak links and take key core technologies into its own hands to avoid being controlled by others.