'The existing solutions and technologies on the market can't meet the requirements of AI's computing power. Baidu decided to develop its own chip.' According to Li Yanhong, 'the computing power of Kunlun chip is compared with the chip originally made with FPGA. The computing power has an improvement of about 30 times, which can be applied to many aspects such as voice, image, and automatic driving. 'But Nandu reporter learned at the scene that the chip is still in trial production, and the specific production schedule is not known.
Coincidentally, mobile phone manufacturers Huawei, Mito, voice recognition manufacturers go out to ask, if Qi Qi are beginning to extend to the upstream chip. After the Sino-US trade dispute, domestic terminal manufacturers have made the chip enthusiasm unprecedentedly. Just like the Internet in the past few years, Many of the chips released this year have also been referred to as 'PPT chips'. Of course, in addition to brand demand, more chips are also seeking a more cost-effective, or more competitive development route.
Why do everyone make chips themselves?
'Baidu has been investing in AI technology on a large scale since about 2011. It has encountered various problems and needs to be solved with innovative methods. One of the important problems is that the demand for computing power has been Improvement. We used the CPU in the early days, and then started to use the GPU, and then use the FPGA to do various optimizations end-to-end. By 2017, we still feel the existing solutions in these markets and The existing technology can not meet our requirements for AI computing power. 'Li Yanhong said.
Coincidentally, in the Nandu reporter interview, many manufacturers of chips have said that the main reason is '(the market) there is no suitable chip'. Go out and ask if Ruqi choose to cooperate with Hangzhou Guoxin to develop chips.
Previously, Yunzhi CEO Huang Wei said that 'speech recognition company does not make its own chip dead end one'. And Ruoqi CEO Zhu Mingming also agrees with this view. 'If someone else does, we use someone else's.' Zhu Mingming said, Everyone knows that chip profit margins are very low. Doing this thing does not make money. 'Sometimes the computing power is not as high as possible, but it needs to be optimized for specific scenarios to find the most cost-effective solution.'
However, according to industry insiders, independent research and development of chips is not to improve the overall computing power, but to optimize some functions. Unlike the CPUs that emphasize complex computing in the past, many artificial intelligence applications, especially 'learning', need to be simple. Repeatedly input a lot, which is why the general-purpose chip such as the 'decathlon' CPU is difficult to perform in the artificial intelligence field.
Going out and asking questions also developed a chip module that asks for information on the front end. 'Answer'. 'For example, after the TV is turned off, the same voice wakes up offline. This solves the important pain points of speech recognition, 'and why other manufacturers have already used it. When a chip front-end signal chip is needed outside the chip, the CEO asked Li Zhifei to explain: 'One is (general chip) expensive, and the other is that the integration effect is not good or convenient. (Going out to ask) 'Question' AI chip The module can be used directly with USB. '
Baidu's 'Kunlun' emphasizes the use of cloud computing power. 'This is China's first full-featured cloud AI chip, which can be applied to many aspects such as voice, image, autopilot, etc. It is also very powerful, and it is Programmable, but also very flexible. 'Li Yanhong also revealed that the advantage of 'Kunlun chip' is 'flexible programmable': 'These chips, together with the software on the chip, the development framework, and various applications constitute a Huge platform and ecosystem'.
According to the Southern Reporter, the 'Kunlun' chip can be adapted to speech recognition, image optimization, automatic driving, and is still in trial production. The production schedule is still unknown.
Is the AI chip so good?
From Yunzhisheng, going out and asking about the release schedule, the chip has only been more than three years from design to mass production, which is far from the general impression that 'a chip industry needs decades of technology precipitation'. But actually There are many kinds of chips, and the production methods and definitions are also different.
After the release of the first artificial smartphone chip Kirin 970 in September last year, Qualcomm related people told the Southern Reporter that the AI unit module is not difficult to build into the chip. Qualcomm can do it before. The so-called AI chip is not independent. A chip, but optimized for some AI functions. '
'It is not necessary to specialize in an AI chip. The key is to release more computing resources and other modules to deal with complex scenes. ' At the beginning of this year, Shang Tang and Qualcomm cooperated in the field of innovative vision and camera-based graphics processing, chip optimization, Shang Tang CEO Xu Li told reporters in Southern China that the development of soft and hard integration is the trend of AI mobile phones. The key is to break the intermediate development layer (SDK) between the low-end of the chip and the terminal application: 'The previous chip layer mainly optimized the operating system, but now AI Requires the chip to directly optimize the application. '
This gave birth to two kinds of AI chips. One is similar to the Mito MT-AI image processing chip, go out and ask 'National Core' such as front-end processing chip, with smaller modules, cheaper, lower power scheme To deal with specific problems, one is to integrate the architecture processing for some specific functions on the general-purpose chip, such as the 'Kunlun chip' released by Baidu.
'Today's chips are not the same as the chips that I have understood in the past,' Zhu Mingming said, today's chips do not need to spend all kinds of IP (intellectual property) cores. Now we are talking about SoC (integrated circuit chip) new products, more It is an optimization at the architectural level, and the combination of IP is optimized according to the demand business. '
For the phenomenon that technology companies are rushing to do AI chips, Wu Gansha, founder and CEO of Detective Technology, told Southern Reporter that 'on the one hand, the threshold of AI chips is lower than that of general-purpose chips. On the other hand, most of them say that they can be made. AI chip, but whether it can be used well or not can be sold in large quantities.