Speaking from Rokid's self-developed AI chip, in the report of Shenzhen Bay last night, we made such a neutral description of KAMINO18:
An AI speech-specific SoC chip that integrates multiple core components such as ARM, NPU, DSP, DDR, DAC, and the size is only equivalent to one-dollar coin. Combined with Rokid phased array technology, CTC model, custom wake-up words, offline speech Instruction, low-power wake-up and other algorithms, the power consumption of the products with KAMINO18 can be reduced by 30%-50% under the whole working condition. It also has better performance in noisy scenarios, and can support more offline functions at the same time. In terms of cost, the overall cost of KAMINO18 can be reduced by more than 30% compared with mainstream mainstream chip solutions.
The Himalayan FM also announced a formal strategic partnership with Rokid. The Himalayan children's smart speaker Xiaoya Mini will soon become an external product using Rokid KAMINO18. At the same time, the capacity of the Himalayan FM in terms of capacity will also be integrated into KAMINO18.
In addition, Rokid CEO Misa also announced that KAMINO18 will support Amazon Alexa, based on language recognition technology, which can seamlessly switch between Ruoqi and Alexa, recognizes that the user calls Ruoqi when speaking Chinese, and calls Alexa when speaking English.
In the post-meeting interview, Misa stated that Rokid is not "selling chips". The chip is only a "competitive element" of Rokid's external output technology solutions. Rokid's product experience and capability in the AI field will still be through technical solutions. Ways to empower partners. In addition, Misa further revealed that KAMINO18 has now won millions of orders.
In spite of the originality of the news, in the report we even shot the pictures as much as possible. This coin-compare chip is still very large and looks like a huge ring-shaped screen on the scene. An AI-capable development board.
And this photo of the spy I actually acquired before the press conference may make the chip look "Small is Beautiful", as Misa said at the press conference:
Half way out to bite gold
June 26th was Rokid’s long scheduled release date. Before that, it was “half-way killing.” Yun Zhisheng and Going Asked had a press conference earlier than Rokid.
At the outreach and interrogation strategy new product launching conference held on May 26th, Li Zhifei made a commercial strategy interpretation of his business for six years. At the same time, he launched "China's first mass-produced AI voice chip module." Probably because of this The name of the module is called “Qinxin.” Going out and asking questions at the press conference and subsequent industry papers are all stressing that this is a “module,” but they are interpreted by the media as unclear as “China’s first AI voice chip."
Whether it is the first model, or whether it is a chip, no one has come out to open the tears. There is no need to ask the officials to go out and ask questions. It is important that Li Zhifei also looks forward to “contributing to the Chinese core”.
According to Li Zhifei, "Qianxin" is based on the intelligent voice interaction SDK launched last year. It integrates the microphone array signal processing technology, voice interaction SDK and customizable semantic skills.
It’s time to introduce Rokid, go out and ask the chip companies behind these stars.
Hangzhou Guoxin Technology is a chip partner of Rokid and asking questions. My understanding of this company began in a press conference held in Shenzhen last October. At that time, the Hangzhou "country" prefixed from the set-top box solution. The old chip company started to organize an event for only 100,000 yuan, invited a partner such as Misa to stand on the platform, and announced the launch of the GX8010 Internet of Things for AI products such as speakers, TV, children's story machine, and car. Artificial intelligence chip.
A brief overview of the characteristics of this chip (for details, please refer to the Shenzhen Bay report - power consumption issues make smart speakers inseparable from fixed power, and the domestic AI chip maker's solution is worth watching):
Local off-line neural network calculation: Built-in gxNPU neural network processor independently developed by Guoxin, compressing calculation weights, making less memory and bandwidth needed for calculation, thus enabling off-line calculation.
Highly integrated: Integrates Cadence Tensilica's highest-order speech DSP Hifi-4, which is designed specifically for intelligent speech DSP can efficiently perform various speech signal processing calculations. At the same time, on this DSP, integrated speech, like Robich, Rokid Algorithmic company's AI algorithm capabilities.
Low cost: Because of the high degree of integration, the peripheral devices required by the product are reduced, thereby reducing BOM costs.
Low power consumption: In the dynamic power consumption, using the advantages of multi-core heterogeneous, rationally arrange the operating frequency and start-stop time of each module; In standby, multi-level wake-up mechanism is adopted.
Who is the cusp of the AI chip?
Regarding the issue of WHO, aside from the news hotspots created by the “Landing Children” as described above, Lingyun, the general manager of the Hangzhou Guomin Artificial Intelligence Division, opened a dialogue with Shenzhen Bay and gave a straightforward answer:
The most eager to do chips is these AI algorithm companies. In the process of empowering and transforming non-intelligent products, they need to better adapt the algorithm's hardware carrier, on the one hand, to facilitate the optimization of the algorithm and improve the efficiency; on the other hand, It is due to better business model considerations.
I used my head of liberal arts students to interpret this point of view and explain it in detail as follows (In this way, those who did not understand the chip, like me at first, might understand it):
AI reformed IoT, but the usage scenarios and equipment types of IoT products are different, and these devices use different chips, which makes it difficult for these AI companies to deal with various chip companies. At the same time, the algorithm capabilities of these AI companies In the process of landing on a hardware terminal (such as a microphone device), it is also necessary to repeatedly perform algorithm optimization on the hardware side, in which more people and objects are required to deal with each other. Uncontrollability, high costs, and low efficiency all contribute to the development of algorithms. AI company wants to make its own special chip idea.
AI companies have to assign their own algorithmic capabilities to all IoT industries. The output of the products is in the form of a software SDK, and the software is more difficult to quote bids at customers because it can be counted as "zero cost." Therefore, AI company considers to output its own capabilities through "hardware". The product upgrade route is generally: software SDK, hardware module, chip.
The chip-based carrier integrates its own cloud service capabilities, AI algorithm capabilities, and the hardware-side algorithm capabilities understood from the perspective of the AI company. The benefits of this are: cost transparency, controllability, high operational efficiency, large-scale future The profit after the birth is considerable, and the prospects for the board can be better planned.
In this regard, AI startups are more flexible and therefore run faster than giants such as BAT.
After becoming a chip company capable of integrating algorithms, what do traditional chip companies think?
"Is it available?"
This kind of disrespectful attitude that is full of "Ha ha," is rather close to the essence of the problem.
The chip business is a “low cost” business. The amount of equipment, cost reduction, and profit growth. This growth curve has become a strategic reference for all chip companies to carry out preliminary product selection and subsequent product sales. .
In the dialogue between Shenzhen Bay and Hangzhou Guoxin, Lingyun described the process of the birth of “self-developed chips.” The custom chip that Guoxin and Rokid cooperated began in the autumn of last year, and the research and development period has been more than half a year. The manpower. The cooperation between Guoxin and the door-to-door question began this year. The main cooperative research and development direction is the front-end voice module.
And we know, "It takes at least 7-8 months or even longer for the chip to go from tape to mass production. Even after the chip is mass-produced, it has to be imported into the customer, and the product is optimized according to the customer's needs. Longer cycle. We all say that the chip threshold is high, this threshold is mainly long cycle. This requires the chip company to have a strong predictive ability, iteration can not be wrong."
For the future, “We are optimistic about this direction of AI. The company also supports the input of human and material resources in this direction.” Ling Yun told us. But because this direction is brand new, it may take some time for the product to reach volume.
The two core chips of Hangzhou Guoxin are 8010 (for smart speakers and children's smart products, cooperation case Rokid), 8008 (for TV set-top boxes, home appliances, cooperation cases go out to ask), when talking about Guoxin will not be in the near future When new products were introduced, Lingyun replied: For the time being, it will be good to have these customer services handy this year.
For geographical reasons, Rokid gave a lot of inspiration to the old company of Hangzhou Guoxin, which brought them one foot into the AI field. This generation of AI chips took two years from the establishment to the mass production, mainly the products. The development cycle is long, and the project decision is actually very fast. When the decision was made two years ago, Guoxin was not sure about the future. But now because of the cooperation with these AI companies, confidence is getting stronger and stronger. Lingyun feels.
As a designer of intelligent hardware devices, how to choose?
In response to this problem, we talked with the team who recently selected a new product for their smart speakers.
Q: As the next team to develop smart products, what are the criteria that you evaluate in choosing which chip to use?
A: The core of chip selection is the trade-off between maturity and cost. Many of the previous speakers used more TV box chips, which are more mature, but the integration is low. The GPU on the box is not used on the speaker. , And at the same time in order to mount multi-channel microphone but also to add additional components, so the cost is generally high. For new chips, especially AI dedicated chip, because it is a special NPU algorithm, so the first look is the algorithm Maturity, followed by the chip vendor's service capabilities, should be able to solve the underlying pit first.
Q: What do you think of the wave of algorithm companies sinking into chip companies?
A: IoT itself is infiltrated for a more certain segmentation scenario. The market prospect of this speaker is very clear, which gives many algorithm companies an opportunity, so deep cooperation with chips or chip companies can enhance their competitiveness.
Of course, we also have to know that traditional chip companies like NVIDIA and Xilinx have strong algorithms and the integration of chips and algorithms will be stronger. Even if the algorithm company does not make chips, the chip company has been doing algorithms.
Looking back, the earliest wave of dual-camera solutions was made by the algorithm company. Basically, Qualcomm and MTK are now equipped with dual-camera algorithms. The future will be the same in other areas. Previously the most powerful dual-camera company, LINX Directly acquired by Apple.
After the Rokid conference was over, a group of old friends surrounded various chats. Inadvertently, one of Rokid’s early investors was told to describe their attitude toward Rokid's spending of nearly 10 million. He said:
Still have to support these dreamers.