AI chips bloom everywhere | technical content is difficult to distinguish

'After the popularization of the AI ​​concept, all parties are looking for business models, expecting AI technology to land as soon as possible, but most of the current AI startups are relying on financing to burn money'

Pre-A round of financing of 340 million yuan, a news of the artificial intelligence field neural network solution company Sugawara technology announced the financing has once again caused the attention of the chip industry.

Ebara Technology was established in Shanghai in March this year. The product is a deep learning high-end chip developed for the cloud data center. It is located in the artificial intelligence training platform. This is Tencent's first investment in the domestic AI chip company, the seed round investor and the capital (Wu Yuefeng) Capital Funds), Zhenge Fund, Datai Capital, Yunhe Capital continue to vote.

In recent years, AI chips are undoubtedly one of the hottest topics. Not only NVIDIA, Google and other international giants have launched new products, Baidu, Ali, etc. have also laid out this field, and the birth of AI chip startup companies such as Cambrian. In the case that the traditional chip fields such as CPU and GPU are far different from the international ones, the Chinese AI chip is expected to achieve cornering overtaking.

AI chips bloom everywhere

From the second half of last year to the first half of this year, many domestic AI start-ups have launched their own chips. Two months ago, Yunzhisheng held a press conference in Beijing to launch its first generation UniOne IoT AI chip and solution. In just two months, there are a number of companies that have released AI chips or modules. Going out to ask questions officially released the AI ​​voice chip module 'Question'; Rokid released the KAMINO18AI voice-specific chip; Spiron also announced that it will be under Launched AI chip in half a year...

According to the usage scenario, the AI ​​chip is mainly divided into the cloud and the terminal chip. At present, the mainstream deep learning artificial neural network algorithm includes two aspects of training and inference. Since the training side needs a lot of data to train the artificial neural network, the training is mainly in the cloud. The cloud is pursuing high performance, the development cost is higher, and the terminal is more focused on low cost and low power consumption. At present, the main layout of Chinese AI startups is here.

In terms of cloud chips, Cambrian officially released the first cloud smart chip MLU100 on May 3 after launching the world's first commercial terminal intelligent processor IP product in 2016. In July, Baidu officially launched at the AI ​​Developer Conference. Kunlun, based on Baidu CPU, GPU, FPGA AI accelerator development. Officials said that this is China's first cloud full-featured AI chip.

According to the global AI chip list released by market research company CompassIntelligence, in addition to NVIDIA, Intel and other traditional chip company giants, Cambrian, Horizon and other AI chip companies are also in the forefront.

Due to its high flexibility, FPGA (Field Editable Gate Array) is considered to be an intermediate solution when the AI ​​algorithm is not mature. The biggest advantage is that the hardware functions of the system can be modified by software like software. Compared with GPU and CPU general-purpose chips, it has higher performance and lower energy consumption.

Shenjian Technology began to purchase FPGAs from Xilinx, put the core algorithm DPU into FPGA, and then sell it to customers in modules, but FPGAs are relatively expensive, and compared with dedicated custom chip ASICs, performance and power consumption. In addition to the FPGA solution, Shenjian Technology is also developing AI-specific chips, which are currently being filmed. A related person in charge of the company told reporters: 'If at this point in time, AI's start-up companies do hardware re-selection. FPGA, maybe a little lagging. '

ASICs are designed for specific purposes. They have high performance and low power consumption, but they are less flexible. They are more suitable for mature and fixed AI algorithms. Once mass production, the cost will be significantly reduced.

Huang Zhi, founder and CEO of Yunzhisheng, said that whether it is CPU or GPU, FPGA, the existing chip architecture is not specifically designed for AI, can not meet the IoT AI computing needs, and considers too much backward compatibility, Therefore, the performance is far from optimal. 'Based on the business aspect of the chip products, the repeated verification of the scene, and the judgment of the final stage of the AIOT (Artificial Intelligence + Internet of Things), Yunzhisheng clearly stated in 2014 that it must independently develop the Internet of Things. The AI ​​chip. 'He said that if the cloud knows not to be a chip, it will die. In this regard, Rokid founder and CEO Zhu Mingming also agreed that the company that makes the voice will always do the chip, 'now the top companies do.'

Going out and asking the founder and CEO Li Zhifei when asked why he wants to make a chip module, mainly to meet specific needs, 'such as smart TV so-called intelligent this year, far-field voice interaction is a strong demand, but on the market There is no good solution. One is expensive, the other is that the effect is not so good, and the integration is not so convenient.

Another AI voice company, Spirit, also announced that it will launch a smart voice chip after it announced the D round of financing of 500 million yuan. It is expected to be released in the second half of the year.

Where is the AI ​​chip difficult?

The chip industry is a high-input, high-risk, slow-return industry. Many industry insiders told reporters that the chip development cycle is very long, usually takes about two years from project to listing. As a startup, especially in algorithms. If you independently develop chips, you will face tremendous pressure in terms of time and money. The most important reason is the high cost of the chip and zero tolerance for errors.

Unlike software, which can be modified and quickly iterated, the iteration cycle of the chip will be very long. If it has been streamed, correcting an error may take several months and then spend millions of dollars to stream again. 'You have to have a very strong psychological quality, Extremely rigorous work style, and it is better to kill a thousand things for anything, not to miss one attitude, not only to have such a person, but also to need such a team to do this well. Chen Zhongmin, vice president of research and development of chip technology, told reporters.

This is the characteristics of the chip industry itself, but the current AI algorithm has not been fixed. If you directly make a dedicated chip, there is no doubt that there is a new risk. Zhang Yongqian, general manager of the Horizon Intelligent Solutions and Chips Division, told reporters that traditional chip companies are designing IP and Before you make a chip, you have already identified the target customer. 'If you make a big decision, you have to have a big head customer to work together. It is equivalent to the chip has not come out, you have already determined who will use it, how to use it It's a very thorough study of a market. But this is the traditional way, the AI ​​chip is different, he pointed out, 'The current AI landing is still early, you can't know who will use you beforehand, This time is a certain risk, but also needs to test a certain vision. If you have to stare at a large number of markets to do AI chips, the first judgment may be wrong, the second time you are late is made. When you see the quantity to do, there are some pre-judging companies that have already made it, waiting in that market. '

Hangzhou Guoxin released its first voice AI chip GX8010 at the end of October last year, which was officially launched at the beginning of this year. Ling Quyun, general manager of Guoxin AI Division, said in an interview that the company decided to lay out AI chips in early 2016. At the time, the chip did not have a clear customer. 'Why did we dare to make this decision? We believe that the underlying architecture of these algorithms is based on neural networks. No matter how your form changes, that core will not change. On the one hand, it's hard to get in touch with customers without products. 'We also talked to customers, but usually, when you don't have one thing, when you talk to customers about the demand, usually don't talk. It is very deep. ' He said that after half a year of listing, the chip has already had a million orders.

It is precisely because the core is not easy, there are AI algorithm companies choose to cooperate with the chip company to serve customers. The above-mentioned Hangzhou Guoxin was mainly engaged in digital TV, home multimedia chip design and system solution development. Going out to ask the chip module, Rokid's chips are all cooperating with the company, and Spirit is also a partner of the company. Lingbiyun told reporters that when working with these AI companies, 'we get out of the chip, they come out with algorithms, let's push customers together'. According to different market scenarios, choose different partners. 'Our cooperation with Rokid is mainly smart speakers. I am going to go out and ask about cooperation. Mainly TV, set-top boxes and some home appliances. The cooperation with Spirit is mainly based on home appliances, IoT. Lord. Because the fields are different, the algorithms also need to be optimized. '

Rokid, a chip manager told the First Financial Reporter, in the cooperation between the two parties, Rokid proposed architecture and performance requirements, Guoxin designed to produce chips and provided the underlying bsp (board-level support package), 'We are responsible for outputting os based on Rokid voice service solution. '

Zhu Mingming said that Rokid is not a chip company, but the chip will become a very competitive element. 'If this competitive element does not exist, we will not make chips.' He pointed out that today's chips are basically SoC, 'SoC There are 90% of things inside, Rokid doesn't have to spend energy on all kinds of IP. Rokid does not use chips as a starting point. Because people in the industry know that chip profits are particularly low. If there is no market, I will do it; If there is on the market, I will use it. '

Huang Wei also said that for Yunzhisheng, core making is not an end, but a starting point.

Industry or more rational

After the popularity of the AI ​​concept, all parties are looking for business models, expecting AI technology to land as soon as possible, but most of the current AI startups are relying on financing to burn money, and AI chips are also considered as a way to land AI technology, but For now, this road is not easy.

Some insiders believe that the AI ​​chip industry will usher in the merger and acquisition period, and let everyone more clearly see the difficulty of making chips.

Taking FPGA-based Xilinx's acquisition of Shenjian Technology as an example, Xilinx said it will continue to increase its investment in Shenjian Technology and continue to promote the company's common goal of deploying machine learning acceleration from cloud-to-end applications. Shenjian Technology is mastering the DPU algorithm, but the chain of the chip is too long, and the DPU is not enough. If you only rely on yourself, you must constantly increase the cost of chip design and R&D in the visible range.

In an interview with the First Financial News reporter, Chen Zhongmin said: 'Why is the chip so difficult? It is not that the knowledge is complicated and the capital investment is high. The more important reason is that from the research and development level, the biggest difference between the chip and other industries is the error. Zero tolerance. '

He pointed out that the cost of a single film is getting higher and higher. If you use the most advanced 7-nanometer process, it will cost hundreds of millions of yuan to stream a film. Therefore, the tolerance for error is almost zero. Even the more mature 40 Nano and 55nm processes, a mask cost also needs millions of dollars, not to mention tens of millions of dollars in design software.

Wei Shaojun, director of Tsinghua University's Microelectronics Institute, pointed out that AI is undoubtedly very important, but the development of AI chips is likely to encounter a setback in the next 2-3 years. Today's part, even most entrepreneurs will become this technological change. 'The martyrs' in the middle.

Zhang Yongqian also told reporters that the AI ​​chip market will definitely be big in the future, but it can't accommodate so many companies, so some companies will die. 'This is also normal. When any new technology comes up, it is especially big like AI. When a low-level technology emerged, there was a bubble. When the Internet bubble burst in 2000, many big Internet companies went bankrupt, laid off employees, and then got up again. Industry has a cycle, now it is already at its highest point, I I feel that the next year will definitely come down, and then return to a rational growth.

2016 GoodChinaBrand | ICP: 12011751 | China Exports