The launch of micro-network news (text / Jiufang Fang), Amazon Echo, Ali Tianmao and other AI smart speakers, forced the upstream voice chip to usher in a new pattern. From May to July, there are more than 5 voice technology entrepreneurs in China. The company released AI voice chips.
On May 16th, Yunzhisheng released the first AI series chip UniOne for the Internet of Things and the first generation chip 'Yuyan'; On May 24th, went out and asked to release its first AI voice chip module, Mobvoi A1. On July 2nd, Rokid released its AI voice-specific SoC chip KAMINO18; at the same time, the company's CEO Gao Shixing confirmed that the company is building an AI voice ASIC chip, which is expected to be streamed in the second half of this year. Voice chip 'three steps 'The future of AI voice has come? The development of voice chips has gone through three stages, general-purpose chips, dedicated chips and AI chips. In the early stage of intelligent voice equipment (2014-2015), due to the long chip development cycle (18-24 months) ), R & D investment is high, in the case of terminal sales is difficult to support the outbreak of chip scale, the market uses universal chips.
General-purpose chip, which refers to the combination of AP chip/tablet chip and other +Codec chip/DSP chip. The Codec chip performs digital-to-analog signal conversion, and DSP processes digital signals, including echo cancellation, noise suppression, and voice noise reduction/enhancement. Etc., finally joined the cloud computing support. The representative chip is MediaTek MT8563 and Quanzhi R16 (tablet chip).
The dedicated chip is the second stage of the development of voice chips. It adopts a CPU suitable for voice processing, plus a multi-channel microphone array interface, and supports echo cancellation, noise suppression, sound source localization, and voice enhancement technologies on the speech algorithm. It has both computing power and low power consumption. The representative chips are MediaTek MT8516, Conexant CX20924, Jingchen Semiconductor A113, Rockchip RK3036 and Beijing Junzheng X1000. These chips do not have a built-in neural network accelerator. Cloud implementation.
Some analysts pointed out that the characteristics of dedicated chips are that speech recognition, semantic understanding, speech synthesis, task execution, etc. are all carried out in the cloud, but there is a problem of voice interaction 'delay' in the cloud. The demand for the network limits the equipment. Use space and bring data and privacy crises.
The AI voice chip, which is the third-generation voice technology, solves these problems: (1) Integrating a dedicated AI processor module (NPU) to accelerate local machine learning algorithms; (2) Voice AI chips are not only integrated CPU, NPU, also integrates DSP signal processing, Wi-Fi/Bluetooth and other modules; (3) It can realize 'end side' intelligence, convert common functions from the cloud to the local, and operate offline and solve user data privacy problems. Intellect launched the CI1006 in 2016, the GX8010 launched by Hangzhou Guoxin at the end of October 2017, which is a typical AI voice chip representative.
The above is the 'three-step' of the development of voice chips. From the current terminal market to the adoption ratio of the above three types of chips, the dedicated voice chip is the leader, with data showing that 70% of the sales of 30 million smart speakers in 2017 were MTK includes. Analysts believe that there are two reasons for the use of dedicated chips. First, the general-purpose chips are outdated. Most of them are borrowed from the flat-panel/OTT AP chip. They simply combine the multimedia digital encoder with the DSP. The effect is not great; the second is that the AI chip that is new is just getting started, and the ecology is still being established.
Rokid vice president and head of the basic platform Zhou Jun said: 'At present, the general-purpose chip has been difficult to meet the needs of smart speaker scenes. Our early products also used a general-purpose chip. The biggest challenge is the real-time wake-up function, which requires two cores. Working at the same time for a long time, high power consumption and not portable, sometimes requiring quad-core or even eight-core computing speed.
At present, MediaTek, Conexant, Jingchen, Ruixinwei, Junzheng, Torch and other manufacturers are the main force of dedicated voice chip shipments, then, with Guoxin, Rodik, go out to ask, Yunzhisheng and more With the emergence of AI Voice, will AI voice chips eventually replace dedicated voice chips, leading the terminal application market?
Ling Yun, general manager of Hangzhou Guoxin Artificial Intelligence Division, told reporters on the micro-network that it is difficult to determine whether the AI voice chip will completely replace the dedicated voice chip. The ultimate goal of the AI chip is to apply the product. Different routes and practices, find the right application scenario.
Zhong Haowei intelligent voice platform leader Lao Yuyuan also told reporters: 'At the beginning of the AI chip, many companies are building their own technical routes, based on the previous accumulation of AI solutions, it is difficult to judge who will eventually win. The key point is that it is not the time to kill, it is necessary for the industry to work together to build this market.
Respondents who did not want to be named said that with the outbreak of intelligent voice terminals, Yunzhisheng, go out to ask questions, Rokid, Spirit and other voice technology processing companies, through the 'customized' with chip companies such as Guoxin The way, added to the array of AI voice chip / module development, although the time lags behind MTK, AMLogic, Junzheng, torch core, etc., but with the advantage of the AI chip itself, it is destined to gain more market support.
According to the micro-grid reporter, the AI chip developed by Guoxin provides digital signal processor DSP, neural network processor NPU and USB/IIS/IIC/UART standard interfaces. Going out, Rokid and other manufacturers do not need IP design. Only architecture integration is required. Most of these integrations are microphone array signal processing, noise reduction, wake-up technology, voiceprint recognition and some voice skills. Although Yunzhisheng is a self-designed uDSP and DeepNet architecture, it is functionally superior to the above two. The chip is basically the same. In short, the three types of voice chips still have their own markets, and the final performance remains to be seen.
Scene custom chip Ten million applications can recover costs
At present, the special scenes have different requirements for AI chips. 'In AI scene applications, only deep chip customization can better realize the functions of AI' has become the consensus of the industry. However, the cost of custom chips is high. A hurdle in front of many manufacturers.
Some people in the industry pointed out that AI chips must have enough computing power to run various speech algorithms on the one hand, and a large number of interfaces to adapt to various scenarios on the other hand, while allowing cost and power consumption to meet mass production. Business requirements. This is a big challenge in itself.
'If the company develops its own AI chip and adopts the 40nm process, then the cost may increase rather than decrease. The chip must share the research and development cost by scale. The 40nm process only costs 10 million yuan and is allocated to 1 million PCS. The number of product units), the average cost per piece is as high as 10 yuan, which does not include more high R&D expenses. 'Industry said.
In the interview with Ji Wei.com, Torch Technology also expressed the same view. The gross profit of the chip itself is very low. Taking a 55nm chip as an example, it takes about several million dollars, and the research and development costs are excluded. Said that only those powerful companies that can get financing can have the ability to do chip customization.
In this regard, Zhu Bin, head of R&D platform R&D, does not agree: 'The use of general-purpose chips for smart devices is a knife for killing chickens. Special needs require special chips to solve the pain points. Custom AI chips are precisely reducing costs, and artificial intelligence hardware is calculating power. 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.
Like Zhu Bin's point of view, Kang Heng, vice president of IoT Business Unit, believes that custom chips are designed to save costs rather than increase costs. 'The profit of TV, air-conditioning and other household appliances is enough to cover the high cost of voice modules. However, the cost of small appliances such as fans and electric lights is limited, and the advantages of the modules are weakened. Customers want to do more smart products and sink to low-end products, but there is no suitable chip in the market. Within the product of the yuan, the general-purpose chip is not cost-effective. After building its own AI chip, Yunzhisheng can open the chip solution of voice AI technology to customers, and have greater initiative in cost and supply cycle.
The above two distinct views come from a completely different starting point between chip companies and algorithm companies. According to the reporter, although custom AI chips are expensive, in order to realize smart terminals closer to AI functions, many manufacturers still start to make custom chips. .
In 2016, Rokid and Hangzhou Guoxin developed KAMINO18 is the representative of customized chips. The customized chips of Spirent will be released in the second half of the year. Coincidentally, according to the report of foreign media Information in March this year, Amazon is also designing custom to support intelligence. Speaker Echo's AI chip, at the time, said that Amazon already has 449 employees with chip expertise and skills.
There is a principle in custom chips, that is, there must be enough quantity to support the cost recovery. As for Rokid, Spirit, how much cost Amazon puts into the chip customization process, this cost depends on how many terminals are sold to recover, currently reporter No detailed information was obtained. However, Hangzhou Guoxin Lingyun said that a chip should reach a break-even point. At least the terminal using this chip should reach tens of millions of meters. If it is customized, it is at least one million.
Rokid co-founder Wang Yude also said that the most important point of custom chip is the quantity, the key profit point of the chip, the volume should reach more than five million.
'This is also the difference between Guoxin's AI chip and Google, NVIDIA AI chip,' Lingyun said, Google, NVIDIA is more in the cloud chip, cloud chip is not sensitive to cost and power consumption, and the size of a single chip Can do a lot, but the end side is different, the end side must start from the application scenario, according to the actual scene to do customization, once the sales of this scene is difficult to support the cost of chip customization, it will lose money.
So, what effective solutions are there in the short term? Lingyun stressed: 'Customizing a chip from scratch is not sensible, and the cost cycle is too long. It is recommended that the chip company define the chip development as much as possible to cover more. Application scenarios, it is also recommended that downstream vendors often communicate with upstream chip companies, allowing chip vendors to take into account customer needs as much as possible in front-end design, so there is no need to pay extra costs.
At present, it is understood that products based on Rokid, Yunzhisheng AI chip and AI module have begun to be put on the market, and some enterprises have already received millions of orders, which is a good sign. Rokid Zhou Jun told Jiji Net reporter: 'At present, Rokid chips and solutions have matured, and have been adopted by Internet companies, such as the children's education market. We are confident that we can customize better chips to support customers' better development.'
As a domestic manufacturer of 'nuclear', the head of Zhongtianwei's intelligent voice platform, Lao Yuyuan, firmly said: 'The hundred-box battle makes AI voice interaction a hot spot, but smart speakers are just the tip of the iceberg, and the Internet of Everything is the ultimate. Goal! A chip can't cover all markets, such as AI speaker chips can't be put into the car. We will stick to our own route, do special AI voice chips and customized solutions.'
All in all, the cost of customizing AI chips is a big problem, but many respondents still agree that the value of custom AI chips will be even greater. It is the trend of the times. As for how to solve the cost problem, it depends on Rokid, Yunzhisheng, thinking. The AI chip customization company represented by Bichi can achieve balance of payments within the predetermined time, thus establishing confidence in the industry.