Yunzhisheng Officially Announced the New Multimodal AI Chip Strategy and Plan

From IVM to swift, yunzhisheng's core making road

As we all know, yunzhisheng has always been a voice technology manufacturer. With the rise of artificial intelligence and edge computing, yunzhisheng began to enter the AI hardware chip scheme (IVM) of the Internet of things in 2014 And began to form mass production and shipment in 2015. Among them, customers in the home field cover almost all domestic first-line home appliance manufacturers such as Gree, Midea, Haier, Changhong, Hisense and vantage.

In addition, it is worth mentioning that according to the data released by Huang Wei, founder and CEO of yunzhisheng at today's press conference, yunzhisheng's voice AI has entered more than 500 hospitals since 2015, of which 80% are class III hospitals (such as Beijing Union Medical College Hospital, Shanghai Huashan Hospital, etc.).

In the process of providing services in-depth scenarios, yunzhisheng officially launched the self-developed AI chip program in 2015 in order to make up for the energy efficiency ratio of the general chip scheme under the given cost and power consumption conditions, as well as the weakness in edge computing power and multimodal AI data processing.

After three years of research and development, on May 16 last year, yunzhisheng took the lead in launching the industry's first voice AI chip for the Internet of things - Unione swift and its system solution. The chip adopts yunzhisheng's independent AI instruction set and has deepnet1.0 and udsp (digital signal processor) with complete independent intellectual property rights It also supports DNN / LSTM / CNN and other deep neural network models, and its performance is more than 50 times higher than that of the general scheme.

In order to further accelerate the implementation of "swift" and system solutions, on September 12 last year, yunzhisheng released the solution of smart speaker and smart home based on "swift", and announced that it would open source the solution In combination, the turnkey solution of software and hardware integration for specific scenarios provided to customers and partners can enable customers to stand at a higher design starting point and create more stable and reliable products in a shorter time cycle at a lower cost. At the same time, the open source solution can also ensure that customers can design other long tail product forms and build more stable products based on the provided AI capabilities It is rich in aiot ecology.

It is reported that at present, the full stack solution based on swift chip has been imported into more than 10 solution providers and partners, including Midea, oaks, Hisense, JD, 360, China Ping An, hard egg technology, etc. related products will be mass produced and listed as early as Q1.

The success of swift chip in the market has also successfully promoted the growth of yunzhisheng's revenue. Although the overall environment in 2018 was not very good, according to the data released by yunzhisheng, yunzhisheng's revenue in 2018 soared three times compared with 2017. The reason for such a huge growth is inseparable from the "blossom and fruit" of "swift".

Huang Wei, founder and CEO of yunzhisheng, also said that the revenue of yunzhisheng in 2018 mainly came from chips and software services.

5g promotes the landing of aiot, and multimodal AI chip is inevitable

Then, as a voice AI technology manufacturer, why did yunzhisheng break through the positioning of the original voice AI manufacturer, further expand to the field of visual AI, and announce the promotion of multimodal AI chips?

Huang Wei, founder and CEO of yunzhisheng

Huang Wei, founder and CEO of yunzhisheng, believes that at present, we are on the verge of 5g explosion. The combination of 5g and artificial intelligence will really promote the landing and implementation of aiot. It can be predicted that there will be a huge amount of multidimensional data (such as voice, image, video, etc.) in the future The demand for centralized processing and edge distributed computing is bound to further challenge the computing power of the underlying AI supporting hardware chip.

At the same time, AI applications in the aiot scenario have a strong demand for end-to-end cloud interaction. A strong cloud will make the end more capable, and a strong end can improve the real-time and effectiveness of data processing, so as to enhance the ability of the cloud. The two need to be closely combined, which requires unified consideration of chip design and cloud architecture. The traditional general solution architecture is based on high real-time The computing power in the highly intelligent scene is limited, and many practical needs such as cost, power consumption and security cannot be balanced. Therefore, a multimodal AI chip with multi-dimensional AI data centralized processing capability will become the only way.

Multimodal evolution of Internet of things AI chip

At the press conference of the first generation Unione chip swift, Li Xiaohan, co-founder of yunzhisheng, once pointed out that Unione is not a chip, but a series of chips, representing yunzhisheng's overall vision for the development strategy of IOT AI chips. At today's yunzhisheng 2019 multimode AI chip strategy press conference, Li Xiaohan once again demonstrated the multimodality of IOT from three aspects The necessity of AI chip.

Li Xiaohan, co-founder of yunzhisheng

Li Xiaohan believes that the AI chips of the current Internet of things product line increasingly reflect three trends:

The first is scenario. Chip design is gradually evolving from the original one-sided pursuit of PPA, that is, power, performance and area, to solving specific problems in a vertical field based on the integration of software and hardware, and even cloud services. The chip itself has become an important part of the whole solution, not the only one;

Secondly, end cloud interaction. In different application scenarios of the Internet of things, massive terminal devices must cooperate with the end cloud in order to realize functional intelligence, that is, to form a dynamic balance between edge computing power and cloud computing power. The proposition of end cloud interaction needs the strong support of AI chip, which further deeply affects the chip design and final delivery;

Third, data multimodality. In the 5g driven smart alliance scenario, the data dimensions contacted by the chip will change from the original simplification to diversification, and the data to be processed by the chip will also change from single mode to multi-mode, which poses new challenges to the design of the chip, especially the artificial intelligence chip of the Internet of things.

Technical layout of yunzhisheng multimodal AI chip

In order to realize the strategic landing of multimodal AI chips, yunzhisheng has accelerated the technical layout and made rapid progress in machine vision.

Among them, the lightweight image signal processor for machine vision can preprocess the sensor pictures in real time at the rate of 30 FPS without relying on external memory, so as to further improve the processing speed and effect of subsequent machine vision processing modules.

yunzhisheng lightweight image signal processor for machine vision

With the help of multimodal technology based on face information analysis, functions such as face / object recognition, expression analysis, labeling and lip movement state tracking can be realized, which can provide more playability and flexibility for product interaction and user experience.

yunzhisheng's face recognition, expression analysis and labeling technology

It is particularly worth mentioning that the release of deepnet2.0, the core IP of yunzhisheng multi-modal artificial intelligence, marks the comprehensive entry of yunzhisheng artificial intelligence processing core from the 1.0 voice era to the 2.0 multi-modal era integrating voice, image and other processing capabilities.

Deepnet2.0 is compatible with many reasoning networks such as LSTM / CNN / RNN / TDNN, and supports reconfigurable computing and Winograd processing.

According to the data released by yunzhisheng, the accuracy of deepnet2.0 based on yunzhisheng has reached 99.80% and 99.47% respectively in face recognition evaluation such as lfw2018 and megaface2018, which is close to the existing best performance.

At present, yunzhisheng deepnet2.0 has been verified on FPGA and will be launched on the new multimodal AI chip dolphin in 2019.

In addition, in terms of industry university research cooperation in image and chip technology, yunzhisheng has also reached in-depth cooperation with ASIC, the only Artificial Intelligence Computing Center of NSFC led by Duke University, and is committed to AI chip algorithm compression and quantization technology, as well as the research on non Feng's new AI chip computing architecture, which will further promote yunzhisheng's multimodal AI chip strategy Tamp the foundation.

Three chips under development are exposed and mass production will be started in 2019

Under the background that the first mass production chip Swift has been introduced by a large number of customers and occupies the first mover advantage in the market, yunzhisheng is more radical in AI chip planning in 2019. It plans to launch three AI chips and start mass production in 2019.

One is swift Lite, a lightweight Internet of things AI chip. It supports microphone matrix, lightweight voice wake-up and lightweight voice recognition.

One is a multi-modal AI chip "dolphin" for smart city. It supports camera, ISP, OD, face ID, microphone matrix, voice wake-up, offline speech recognition and synthesis, etc.

There is also a multimodal AI chip "Snow Leopard" for smart travel, in which the latter two integrate the capabilities of voice AI and visual AI at the same time.

According to the data released by yunzhisheng at the meeting, the multimodal AI chip "dolphin" will adopt the 28nm process, and the benchmark is Mobileye eq4. According to the plan, the maximum configurable computing power of yunzhisheng's "dolphin" is 4TOPS, which is 1.6 times that of eq4, and the power consumption is 4.3W, slightly higher than eq4.

In addition, according to Li Xiaohan, the multimodal car gauge AI chip leopard for smart travel scenarios is jointly built with Yikatong technology, an ecological chain enterprise of Geely Group, which also means that once the chip is successfully mass produced, it may have the opportunity to realize rapid landing commercialization with the support of Geely Group.

Summary:

At the end of 2017, baidu took the lead in announcing the free opening of its voice AI capabilities, and then iFLYTEK's iFLYTEK open platform also opened its basic capabilities for free. This also made the voice AI market extremely competitive. Against this background, many voice AI start-ups have placed their hopes on creating ecology by launching voice AI chips, so as to realize the improvement of value.

For yunzhisheng, from voice technology service providers to further expand voice AI chips to the field of visual AI, it is obviously hoped to get rid of the fierce competition in the voice AI market and share the cake of the larger visual AI market.

However, compared with the voice AI market, although the visual AI market is broader, it is also highly competitive. There are not only many unicorns such as Shangtang technology, Kuangshi technology, Yitu technology and Yuncong technology, but also old security manufacturers such as Haikang and Dahua, as well as a large number of small and medium-sized visual AI start-ups, and many visual AI manufacturers have launched or are in the process of launching In active preparation for the launch of visual AI chips / modules.

In addition, for the largest and most valuable market of visual Ai - security market, its barriers are relatively high, and it needs a long period of continuous cultivation, and this market is also the focus of everyone's competition. However, due to the different policies, preferences and requirements of local governments, this market also shows a fragmented trend. Therefore, new players rely on it