'AI Chip' | Still in the era of grasshoppers, running wildly | 'bubble' |

Artificial intelligence is changing all walks of life, and chips are the carrier of artificial intelligence.

In 2018, capital's enthusiasm for semiconductor chips was completely ignited by AI technology. Whether it is a giant company or a startup company, a traditional manufacturing company or an Internet company, the enthusiasm for chips is high. From May to July, Yun Zhisheng, go out and ask, Rokid, Baidu has released AI chips or chip modules, and Spirit confirmed that it is building AI voice chips. Shenjian Technology announced that its AI chips will be available in the second half of this year. Yun Zhisheng founder and CEO Huang Wei even used 'No. Do the chip, mortal' to express your determination to make AI chips.

However, 'until now, there has not been a real AI chip in the world, because the real artificial intelligence is far from being realized.' Zhou Bin, general manager of Heterogeneous Intelligence China, represents the views of some insiders.

"IT Times" reporter interviewed a large number of people in the AI ​​industry and found that for the concept of AI chip, there is no unified consensus at present, and even some investors believe that there is a bubble in the field of AI chips, and most startup companies will disappear.

But in any case, in the age of AI chips in the grass-roots era, there will always be people who have made a " bloody road" for the localization of China's core technology.

Different definitions of AI chips

'Start-ups are probably not worth the AI ​​chip. ' Li Yangyuan, the founder of the chip company Suzhou Mindray Microsystems, has been immersed in the industry for more than a decade. He told the IT Times reporter that a chip from 0 to 1 would take more than 10 years. Cycle, his own chip products are also officially commercialized after 5 times of streaming, and can be placed in the top position in the industry 20 times. If the company develops its own AI chip and adopts 40nm (nano) process, the cost may rise. It is not a reduction. The chip must share the R&D cost by scale. The 40nm process only costs up to 10 million yuan, and allocates it to 1 million PCS (a certain number of product units). The average cost per piece is as high as 10 yuan, which does not include a higher amount. R&D expenses. '

However, Zhu Bin, the head of Rokid's platform research and development, is exactly the opposite of Li Yangyuan. 'Smart devices use general-purpose chips to kill chickens. Special needs require special chips to solve the pain points. Custom AI chips are precisely reducing costs. Artificial intelligence hardware is calculating. There is demand, the low-end general-purpose chip is not enough, and the high-end general-purpose chip has many redundant designs, resulting in high power consumption.

AI chips are also known as AI accelerators or computing cards. In general, AI chips are modules that are used to handle a large number of computing tasks in artificial intelligence applications (other non-computing tasks are still handled by the CPU). Currently, they are mainly divided into GPUs. , FPGA, ASIC and other types, unlike general-purpose chips such as Qualcomm Xiaolong, AI chips are mainly used to handle special tasks, such as the identification of high-definition video in security, data calculation during automatic driving, and so on.

On June 26th, Rokid released a KAMINO18 self-developed AI voice dedicated chip that can support smart speakers and children's story machines. It is different from traditional general-purpose chips such as Intel and Qualcomm. This is an SoC (system-on-a-chip) dedicated to AI voice. The chip, which inherits several core components such as ARM, NPU, DSP, DDR, DAC, etc., is about the same size as a one-dollar coin. Zhu Bin used to design the hardware on the wake-up of the voice, and found many functions in the general-purpose chip. Not enough, but the cost of the attachment still exists. So, Rokid began to customize its own chip in September last year, using DSA (Domain specific architecture) architecture, from the product and algorithm requirements, integrated into the chip by means of heterogeneous computing. In the whole working state, the power consumption of the product can be reduced by 30% to 50%.

The different views of Li Yangyuan and Zhu Bin represent the two different perceptions of AI chips in the traditional semiconductor industry and Internet startups. The people born in the semiconductor circle value the breakthrough of the chip from 0 to 1, and 'from soft to hard' Internet entrepreneurs are more eager to become directly on the basis of 1, 2, 3, 4, through algorithms and software design. From the perspective of the currently released AI chips, their main purpose is to Power consumption to speed up some kind of machine learning algorithms. For example, when used on edge-side and terminal-side devices, it requires extremely low power consumption and extremely high matrix/floating point computing power, which is difficult for general-purpose chips. .

It can be proved by the explanation of the engineer of another leading chip manufacturer. In his view, most AI voice chips on the market belong to a chip that serves 'proprietary' functions (similar to DSP data signal processing). Because it focuses on one or several functions and is limited to a specific scenario, the complexity of design and production is lower than that of a general-purpose chip. In addition, such chips rarely involve technical authorization, and it is easier to do. The AI ​​chip does not start from the bottom layer, but directly from the SoC or general-purpose processor plug-in architecture optimization such as 'coprocessor', through the combination of various IP to accelerate the scene of the upper application, such as language, image.

Heterogeneous

At present, there is no clear definition of AI chips, so how to calculate 'real' is not a good measure. Not long ago, at the 2018 Moving Point International Summit (Hangzhou), Zhou Bin, general manager of Silicon Valley startup heterogeneous smart China, tried to give AI chips. Under the definition, 'the core algorithm of artificial intelligence can be completed with high efficiency and high performance. Since the mainstream algorithm is deep learning now, it means that AI chips must have very good support for deep learning.' From the data point of view, Zhou Bin believes that the computing power of AI chips must exceed 5 trillion times per second, because only by achieving such performance indicators, many specific application calculation results can be comparable to human capabilities.

Zhou Bin's name, 'heterogeneous', is essentially the most direct interpretation of the Internet background of AI chip entrepreneurs. Heterogeneous, as the name implies, is composed of different sources, and the Internet is a typical heterogeneous network. Later Evolved heterogeneous computing, a special form of parallel and distributed computing, often used to coordinate different hardware to meet different computing needs, and to enable code (or code segments) to perform in the most overall performance .

At present, AI chips basically perform heterogeneous computing through a variety of chips. In the past, traditional chip companies only focused on a few types of chips, but now chip companies are beginning to focus on horizontal development, integrating different types of chips, for example, mobile phone SoCs in traditional CPUs. In addition to the (central processing unit), GPU (graphics processor), ISP (online programming), there are also additional processing cores such as NPU (embedded neural network processor) to accelerate AI. 'There are some parts of the heterogeneous chip. It is a universal function. 'Cheng Zhisheng co-founder Kang Heng told the IT Times reporter.

Cost-determining path

An interesting phenomenon is that Internet-born AI entrepreneurs are running into the AI ​​chip hardware field, while traditional chip vendors are using algorithms such as 'soft' to implement AI.

AI technology has three major elements, algorithms, computing power and data. From the perspective of international AI technology, the research and development of algorithm models such as deep learning is not mature, and new algorithm models such as migration learning and capsule networks are developing rapidly in synchronization, AI chips. The method and principle of using this method is still in the exploration stage. In fact, the current mainstream chip manufacturers have not introduced AI chips, and many AI functions are completed by general-purpose chips plus special algorithms and software.

Qualcomm's artificial intelligence engine (AI Engine), which was launched earlier this year, includes both hardware and software. It is equipped with a Neural Processing Engine (NPE) on Qualcomm's core hardware architecture (CPU, GPU, VPS vector processor). NN API, Hexagon neural network library and other software, make artificial intelligence on the terminal side (such as smart phone) application faster and more efficient. Qualcomm's chip products 骁龙 845, 骁龙835, 骁龙820, 骁龙660 support AI Engine, and many domestic mobile phones with AI flag also basically adopt Qualcomm solution, and the face recognition function is better realized by the AI ​​Engine.

But for the current AI chip entrepreneurs and smart home manufacturers, the general-purpose chip is 'too expensive'.

Kang Heng told the IT Times reporter that TV, air-conditioning and other household appliances have a profit that covers the high cost of voice modules, but the cost of small appliances such as fans and lights is greatly limited, and the advantages of the modules are weakened. Do more smart products, sink to low-end products, but the market can not find the right chip, a product within 100 yuan, the general-purpose chip is not cost-effective. ' After building their own AI chip, Yun Zhisheng The chip solution of voice AI technology can be opened to customers, with greater initiative in cost and supply cycle.

However, Li Yangyuan believes that the key technologies of artificial intelligence are different in different stages. 'The processor is not the key technology of artificial intelligence. The dedicated processor only enhances the competitiveness in some working segments. 'He thinks that the sensing segment is sensor-centric. The existence value of the processor is not high; the cognitive segment, the learning segment and the decision segment, too much emphasis on the processor but affect the cost, including the one-time cost and the power consumption cost.

'No part of the premise of a dedicated artificial intelligence chip, the development of artificial intelligence chips should be regarded as an independent matter, rather than a natural extension of software research. 'Li Yangyuan uses human analogy, the sensor is the human body, the brain depends on The algorithm wins.

Li Shoupeng, a semiconductor analyst, also believes that artificial intelligence relies on algorithms, and chips are only carriers. If you want to use ASIC (full custom chip) to do better speech recognition processing, the hardware difference is not big. For speech recognition, recognition Statements are more related to software, network, and training, and the existing cloud and end data exchange delay problems will come along with 5G.

First stop on the ground: Intelligent audio

R&D AI chips are simpler than high-end general-purpose chips and are widely recognized by the AI ​​circle.

Li Zhihan, co-founder of Yunzhisheng and vice president of IoT Business Unit, believes that after decades of development, the chip industry has precipitated a lot of modular things. For example, Qualcomm and MediaTek are ARM-based architecture design chips. Therefore, not every An AI chip must start from scratch, and can use mature modules and products in the industry. However, the core acceleration module of the AI ​​chip needs to be designed from the bottom. From 2014, the cloud, the end, the core strategy will be established, and the R&D will be officially established in 2015. The team, and then in 2018 launched the AI ​​chip 'Swift', Yun Zhisheng spent a full four years, and gradually realized that AI can not only be in the cloud, to land.

In 2018, it was called the AI ​​chip landing year. The so-called landing, refers to the AI ​​chip to be put on the terminal for commercial use. In 2017, China's investment in the chip field exceeded 150 billion yuan. Starting in 2018, these investment industries will continue to be intensive. Landing. The founder of Xinli Capital Group, Chairman Wang Chaoyong recently pointed out that China’s makers spent more than 300 billion US dollars each year to import various types of chips, consuming one-third of the world’s chips, self-sufficiency. The rate is less than 10%. Therefore, the investment in AI chip is the top priority. However, due to the high cost of film and R&D of security AI chips, large-scale shipments have not yet been formed to offset the cost; Mass production, autopilot AI chip security is not up to standard; other specific areas of AI chip overall downstream demand is insufficient, supply exceeds demand, 'the current AI field has a large bubble.'

According to Moore's Law, the performance of the chip will double every 18 months, and the cost will be reduced by half. But the fundamental secret of making money in the semiconductor industry is still large-scale. Can it have enough large shipments and commercial market, the AI ​​chip is smooth? The key to 'landing' on paper. Earlier media reports said that even in the security field, 'Dayu' Hikvision, the annual demand for Nvidia is only 200,000.

From this point of view, intelligent audio may be the earliest market for AI chips. Research firm Canalys Research (hereinafter referred to as Canalys) released a report that by the end of this year, the number of smart audio will reach 100 million, almost last year. 2.5 times. Last year, the number of intelligent audio is less than 50 million. In the next few years, the number of smart audio will continue to grow, and by 2020 its holdings will more than double, reaching 225 million.

In May of this year, Yunzhisheng released the AI ​​chip UniOne for the Internet of Things, which is used for edge calculation in the terminal. It can provide service solutions for smart audio, smart home, smart home appliances, etc. The AI ​​voice chip module released for questioning 'Question' has been mass-produced, customers can place orders.

Li Zhifei, the CEO who went out to ask, believes that the chip is a long-cycle industry. From the concept, it must go through system design, module design, simulation verification, line synthesis, place and route, flow film production, package testing, driver development, solution adaptation. After a very long process, once the chip is made difficult to modify it like software, it must be redesigned, the iteration cycle is long, and the cost is high. The chip itself is a computing hardware carrier, and different chip adaptations are different. Algorithms and application scenarios, AI chips must have enough computing power to run various voice AI algorithms on the one hand, and a large number of interfaces for various scenarios on the other hand, while allowing cost and power consumption to meet large-scale quantities. Commercial requirements for production.

'Can be industrialized and can pose a competitive threat to foreign similar products in the market is the standard for a successful chip.' An industry insider said. From this point of view, China's AI chip has just started, Jianghu Cao, The bubble is going to break.

2016 GoodChinaBrand | ICP: 12011751 | China Exports