Baidu helps AI chips transform technology

According to Jiwei.com, Baidu's AI chip division Kunlun recently completed a round of financing. People familiar with the matter revealed that Kunlun's valuation is about US$2 billion. Currently, Kunlun chips are mainly used by Baidu for smart electric vehicles and cloud computing. Baidu will continue to make efforts in AI chip-related fields.

There are many complex calculations in artificial intelligence algorithms. Most current AI processors use a single instruction to call dedicated complex computing units to implement complex calculations. Since these complex computing units occupy a relatively large logical area, in a multi-core AI processor, if each processor core has exclusive use of these complex computing units, it will occupy a large chip area and the implementation cost is too high; and in actual application scenarios, the frequency of use of these complex computing instructions is not particularly high. When each processor core has exclusive use, the utilization rate of these complex computing units is not high.

To this end, Baidu applied for an invention patent entitled "Complex computing devices, methods, artificial intelligence chips and electronic devices" (application number: 202010612288.3) on June 30, 2020, and the applicant was Beijing Baidu Netcom Technology Co., Ltd.

Figure 1 Schematic diagram of AI chip structure

FIG1 is a schematic diagram of the AI ​​chip structure proposed in the present invention. In the AI ​​chip, multiple instruction sources 100 are connected to a complex computing device 200, and the instruction source 100 may be an AI processor core. The complex computing device 200 includes an input interface 210, multiple computing components 220, and an output interface 230. The input interface 210 may be connected between multiple AI processor cores 100 and multiple computing components 220, and the output interface may be connected between multiple computing components 220 and multiple AI processor cores 100.

A plurality of computing components 220 constitute a special function unit, each computing component 220 has independent computing capability and can implement a certain type of complex computing. The computing component 220 includes at least one of the following: a dedicated integrated circuit, a chip, and a field programmable gate array.

The AI ​​processor core 100 can decode the instructions to be executed when receiving the instructions to be executed, and splice the decoded data into complex computing instructions. The AI ​​processor core 100 can send an instruction request to the input interface 210 of the complex computing device 200, and the input interface 210 then obtains the complex computing instructions from each AI processor core 100, and arbitrates them to the corresponding computing components 220 according to the computing type in each complex computing instruction.

The computing unit 220 obtains source operands from the received complex computing instruction to perform complex computing, and generates a computing result instruction to feed back to the output interface 230. The computing result instruction includes an instruction source identifier, a computing result, and a write-back address, etc. The computing result is the computing result of the computing unit 220 performing complex computing on the source operands, and the instruction source identifier and the write-back address come from the data in the complex computing instruction received by the computing unit 220.

The output interface 230 receives the calculation result instruction from each calculation component 220, and arbitrates the calculation result and write-back address in each calculation result instruction to the corresponding instruction source, such as the AI ​​processor core 100, according to the instruction source identifier. The AI ​​processor core 100 writes the calculation result into the internal register according to the write-back address.

In short, Baidu's AI chip patent, through weak coupling with each instruction source, realizes the sharing of SFU by multiple instruction sources, and reduces the data path when the instruction source calls SFU for complex calculations, thereby reducing the area overhead and power consumption overhead of the AI ​​chip.

Baidu is a world-leading AI company that has established leading advantages in AI computing power, algorithms, open platforms, and developer ecosystems, and has entered a fast lane of strong growth. Under the general trend of the integrated development of cloud, AI, and the Internet, Baidu has laid out a new multi-engine growth pattern and accumulated strong potential to support future development.