The field of AI chips in Huicong Security News has been hot in recent years. Whether it is the capital market or the mass entrepreneurship, it has shown great enthusiasm for it. What kind of needs and thinking does Yunzhisheng have for AI chips? Next, this article will be from the perspective of AI and IoT integration (AIoT) for your detailed analysis.
1. Market opportunities for human-computer interaction in the AIOT field
Since 2017, the word "AIoT" has been frequently screened and become an industry hotspot for the Internet of Things. “AIoT†means “AI+IoTâ€, which refers to the integration of artificial intelligence technology and Internet of Things in practical applications. At present, more and more people have combined AI and IoT. As the best channel for intelligent upgrade of major traditional industries, AIOT has become an inevitable trend of the development of the Internet of Things.
In the market based on IoT technology, scenes related to people (such as smart home, autonomous driving, smart medical, smart office) are becoming more and more. As long as it is a place where people are connected, it is bound to involve the needs of human-computer interaction. Human-computer interaction refers to the process of information exchange between a person and a computing machine using a certain dialogue language between a person and a computer in a certain interaction manner. The range of human-computer interaction is very wide, ranging from light switches to dashboards on airplanes or control rooms in power plants. With the outbreak of smart terminal devices, users have also put forward new requirements for the interaction between people and machines, which makes the AIOT human-computer interaction market gradually stimulated.
Figure: AIOT development path
Taking the smart home market as an example, the data shows that the scale of China's smart homes will reach 180 billion yuan in 2018. By 2020, the smart home market will reach 357.6 billion yuan. Analysts predict that the global smart home market will reach more than 500 billion yuan in 2021. The human-computer interaction needs and prospects of the AioT market in the rapid outbreak are undoubtedly expected.
The digitization process of human life has lasted for about 30 years. In these years, we have experienced the evolution from the analog era to the era of PC interconnection and the era of mobile internet. At present, we are in the process of evolution towards the Internet of Things era. From the perspective of interaction, we can see that machines are becoming more and more "accommodating" people: from the keyboard and mouse of the PC era to the touch screen of the mobile era, NFC and various MEMS sensors, and then to the Internet of Things era is booming. The interaction of voice/images, usage thresholds is getting lower and lower, which has led to the involvement of more and more users. At the same time, we need to notice another profound change, that is, due to the evolution of interaction (at least one of the important reasons), a large number of new dimensions of data are constantly being created and digitized, such as work materials and entertainment in the PC era. Programs, users of the smartphone era use habits, location, credit and currency, and then to the possibilities of the Internet of Things.
In the Internet of Things era, the interaction mode is developing in the direction of ontology interaction. The so-called "ontology interaction" refers to the basic way of interaction between people, such as voice, vision, movement, touch, and even taste, starting from the human body. For example, control the home appliance by sound, or the air conditioner uses infrared to determine whether it should cool down, and combine the voice and infrared to control the temperature (when no one is in the room, even if the TV program mentions "cooling down", the air conditioner also Do not react).
The new data is AI's new nourishment, and a large number of new dimensions of data are creating endless possibilities for AIOT.
From the perspective of the development path of AIOT, the current industry generally believes that it will experience three major stages: stand-alone intelligence, connected intelligence and active intelligence.
Stand-alone intelligence refers to the fact that smart devices wait for users to initiate interactions, and devices and devices do not interact with each other in this process. In this scenario, a stand-alone system needs to accurately perceive, recognize, and understand various types of instructions of the user, such as voice, gestures, etc., and make correct decisions, executions, and feedbacks. The AIoT industry is at this stage. Take the home appliance industry as an example, the past home appliances is a functional machine era, just like the previous mobile phone button type, to help you lower the temperature and help you achieve food refrigeration; now the home appliances realize a single machine intelligence, that is, voice or mobile phone The remote control of the APP is used to adjust the temperature, turn on the fan, and the like.
Unable to connect and interoperate smart items, just an island of data and services, far from meeting people's needs. In order to achieve the continuous upgrading and optimization of the intelligent scene experience, the first thing that needs to be broken is the island effect of single product intelligence. The interconnected intelligent scene essentially refers to a matrix of products that are interconnected. Therefore, the mode of “one brain (cloud or central control) and multiple terminals (perceptors) becomes inevitable. For example, when the user in the bedroom says to the air conditioner to close the curtain of the living room, and the air conditioner and the smart speaker in the living room are connected, they can negotiate and make decisions with each other, thereby making the action of closing the living room curtain by the speaker; or When the user speaks "sleep mode" to the air conditioner in the bedroom at night, not only the air conditioner automatically adjusts to the temperature suitable for sleep, but also the TV, the speaker, and the curtains and lamp equipment in the living room automatically enter the closed state. This is a typical scene of interconnected intelligence through the cloud brain, with multiple perceptrons.
Active intelligence refers to the intelligent system is always on standby according to user behavior preferences, user portraits, environment and other information. It has self-learning, self-adaptation and self-improvement capabilities, and can actively provide services suitable for users without waiting for users to make demands. Just like a private secretary. Imagine such a scene. In the early morning, with the change of light, the curtains automatically open slowly, and the speakers send a soothing wake-up music, and the fresh air system and air conditioner start working. You start to wash, and the personal assistant in front of the wash station automatically broadcasts today's weather, dressing advice, etc. After washing, breakfast and coffee are ready. When you walk out of the house, your home appliance automatically loses power and waits for you to turn it back on when you go home.
Second, AIOT's human-computer interaction needs for AI chips
Edge computing refers to an open platform that integrates network, computing, storage, and application core capabilities on the edge of the network near the source or data source. It provides edge intelligent services to meet industry digitalization in agile connections, real-time services, data optimization, and application intelligence. Key requirements for security and privacy protection. There is a very analogy in the industry. Edge computing is like the nerve endings of the human body. It can handle simple stimuli and feed back the characteristic information to the cloud brain. With the implementation of the AIOT, in the scenario of the Realm of All Things, devices and devices will be interconnected to form a new ecosystem of data interaction and sharing. In this process, the terminal not only needs to have more efficient computing power, but in most scenarios, it must also have local self-determination and responsiveness. Take smart speaker example, which not only needs to support local wake-up ability, but also has the ability to reduce noise. Because of real-time and data validity, this calculation must happen on the device side instead of the cloud.
As the most important landing scene of AIOT human-computer interaction, the smart home industry is attracting more and more enterprises to enter. Among them, there are technology giants such as Apple, Google, Amazon, etc., as well as traditional home appliance manufacturers like Haier and Samsung. Of course, there are also Internet upstarts such as Xiaomi and Jingdong. Based on the idea of ​​interconnected intelligence, in the future AIOT era, each device needs to have certain perception (such as pre-processing), inference and decision-making functions. Therefore, each device side needs to have independent computing power independent of the cloud, namely the edge calculation mentioned above.
In the smart home scenario, interaction with terminal devices through natural voice has become the mainstream in the industry. Due to the particularity of the family scene, the home terminal equipment needs to accurately distinguish and extract the correct user commands (instead of the invalid keywords that the family members did not intend to say during the conversation), as well as the sound source, voiceprint and other information. Therefore, in the field of smart home Voice interaction also puts higher requirements on edge calculation, which are shown in the following aspects:
1) Far from noise reduction, wake up
In the home environment, the sound field is complex, such as TV sound, multi-person dialogue, child play, space reverberation (kitchen cooking, washing machine and other equipment work noise), these sounds that easily interfere with the normal interaction between the user and the device, the probability will be the same Time exists, which requires processing and suppression of various disturbances, making the sound from real users more prominent. In the process of this process, the device needs more information to assist in the judgment. An essential function of voice interaction in home scenes is to use a microphone array for multi-channel simultaneous sound recording. By analyzing the acoustic space scene, the spatial positioning of the sound is more accurate and the voice quality is greatly improved. Another important function is to help distinguish the real user through the voiceprint information, so that his voice is more clearly distinguished from the harassment of many people. These need to be implemented on the device side, and require a lot of computing power.
2) Local identification
The local identification of human-computer interaction in the home field is inseparable from the edge calculation, which reflects two aspects:
High frequency words. From the actual statistics, the number of common keyword instructions of users in a specific scenario is limited. For example, car products, the most commonly used by users may be "previous/next". The most common command for air-conditioning products is "on/off". The words frequently used by these users are called high-frequency words. . For the processing of high-frequency words, it can be placed locally without relying on the delay of the cloud, thus giving users the best experience.
Networking rate. In the process of smart home products, especially home appliances, the networking rate is a problem. How to make users feel the power of voice AI without networking, and to cultivate users is also an important role of edge computing.
3) Balance of local/cloud efficiency
In the natural language interaction process in the home field, when all the calculations are put into the cloud, the part of the acoustic calculation will put a lot of pressure on the cloud computing, which will cause a large increase in the cost of the cloud platform; on the other hand, the calculation delay, Damage the user experience. Natural speech interaction is divided into two parts, Acoustic and Natural Language Understanding (NLP). From another dimension, it can be seen as part of "business-independent" (voice-to-text/acoustic calculation) and "business-related" (NLP). Business-related parts need to be solved in the cloud, such as users asking for weather, listening to music, etc., then the device's understanding of user statements and the acquisition of weather information must be done through networking. However, for user speech-to-text conversion, such as issuing instructions "turning on air conditioning, increasing temperature, etc.", some or even most of the calculations are possible locally. In this case, the data uploaded from the local to the cloud will no longer be the compressed voice itself, but the more streamlined intermediate results or even the text itself, the data is more streamlined, the cloud computing is simpler, and the response is more For the sake of speed.
4) Multimodal requirements
The so-called multi-modal interaction is the interaction of multiple ontology interactions, such as combining various senses, such as text, voice, vision, action, environment, and so on. People are a typical example of multimodal interaction. In the process of communication between people, expressions, gestures, hugs, touches, and even smells play an irreplaceable role in the process of information exchange. Obviously, the human-computer interaction of smart homes is bound to be more than a voice modality, but requires multi-modal interaction and parallelism. For example, if a smart speaker sees that someone is not at home, then there is no need to respond to the wake-up words that are mistakenly released on the TV, or even to adjust to sleep; if a robot feels that the owner is watching him, then Will actively greet the owner and ask if you need help. Multi-modal processing undoubtedly requires the introduction of common analysis and calculations for multiple types of sensor data, including both one-dimensional speech data and two-dimensional data such as camera images and thermal sensing images. The processing of these data does not require the ability of local AI, and it also imposes a strong demand on edge computing.
Third, the AI ​​chip demand brought by edge computing
The AI ​​algorithm puts forward higher requirements on the parallel computing power and memory bandwidth of the device-side chip. Although the traditional GPU-based chip can implement the inference algorithm in the terminal, the disadvantages of high power consumption and low cost performance cannot be ignored. In the context of AIOT, IoT devices are given AI capabilities to complete AI operations (edge ​​operations) while ensuring low power consumption and low cost. On the other hand, IoT devices are different from mobile phones in various forms, requiring fragmentation. Seriously, the demand for AI computing power is not the same, so it is difficult to give a general-purpose chip architecture across devices. Therefore, only from the IoT scenario, designing a customized chip architecture can greatly improve performance while reducing power consumption and cost, while meeting the needs of AI computing power and cross-device form. Compared with traditional chips, customized AI chips have absolute advantages in terms of computational efficiency and memory bandwidth, and their advantages are mainly reflected in the following points:
First, parallel computing architecture and dedicated matrix accelerator technology, such as SystolicArray architecture or more complex parallel computing architecture to achieve the utilization of computing units, even using a specific matrix accelerator such as Winograd, reduce the computational complexity of matrix operations, thereby improving computational efficiency. .
Secondly, from the perspective of reducing the bandwidth of the external memory, the memory bandwidth is reduced by data compression or pipeline technology between related functional modules. Taking NVIDIA's open source AI engine NVDLA as an example, a dedicated data path is designed between modules such as convolutions, activations, and pooling. The data interaction between modules is not through the system memory, but by a dedicated data channel.
Figure: NVDLA core architecture
As the efficiency of the chip increases and the bandwidth of the external memory data decreases, the number of clock cycles and memory accesses required for chip operations will be greatly reduced. Therefore, compared with the general-purpose chip, the AI ​​chip can complete the calculation of the same amount of tasks in machine learning at a lower main frequency and a smaller chip area. Using a lower frequency, not only reduces the dynamic power consumption of the chip, but also reduces the operating voltage of the chip, thereby further reducing the dynamic power consumption of the chip. In addition, the low main frequency makes more choices in the chip processing process, further affecting the static power consumption of the chip.
Since the chip area and power consumption directly affect the choice of chip package, the AI ​​chip has an advantage over the conventional chip in the chip package. From this point of view, the price/performance ratio of AI chips will be much larger than that of traditional chips.
Fourth, the challenge of AI chip landing
The AI ​​chip combines flexibility and versatility while ensuring high performance and high energy efficiency. The AI ​​algorithm has a wide application field, a variety of algorithms and a fast evolution of the algorithm, so it puts forward higher requirements for the AI ​​chip architecture design. Only the performance and versatility of the AI ​​chip will have a broad market and a long life cycle. Only when the high degree of coupling between the chip architecture and the software algorithm is formed, the cost performance of the chip can reach a high value, so it is difficult to design an efficient AI chip without leaving a deep understanding of the algorithm. The direct hardwareization of the AI ​​algorithm will greatly reduce the flexibility of the chip, so the hardware acceleration of the AI ​​chip is often reflected in the lower mathematics than the algorithm. Since the AI ​​algorithm is built on a series of mathematical operations, designing a cost-effective and flexible chip requires data dependencies between mathematical operations and operations. On the basis of a large amount of data statistics, based on the complexity of the data operation, frequency of occurrence, data dependence and other information, the basic mathematical operation instructions and data handling instructions are refined, and the definition and implementation of the chip architecture are completed according to these instructions. Therefore, AI chip architecture design and implementation is a small link in the middle of AI chip design, and AI instruction set design is the more critical factor for the success of AI chip.
Doing AI chips is not an arms race. Any product has its product positioning in the process of being introduced to the market. AI chips are no exception. The specific AI chip is oriented to a specific scenario, and the scenario is determined by the product and market that the chip is facing. The different scenarios have different requirements for the AI ​​chip in terms of price, power consumption, and supported functions. For example, large service robots and smart switches may have a large contrast to the price requirements of AI chips. After all, the prices of the two products are very different. For the same price of AI chips, they are expressed in terms of product cost. The acceptance may be diametrically opposed. Therefore, the AI ​​chip not only has to be done, but also has to be sold.
In the AI ​​software ecology, the AI ​​development framework (Framework) is relatively fragmented, such as TensorFlow (Google), CNTK (Microsoft) and Torch7 (Facebook), etc., the entire industry has not yet formed a unified standard. The exploration of standardization mainly includes ONNX jointly launched by Microsoft and Facebook, and the neural network exchange layer standard such as NNEF launched by Khronos. It is undeniable that the neural network switching layer standard is a shortcut to solve the current fragmented AI framework, but the standard is in its infancy, and the maturity of the standard still needs a long way to go. Therefore, AI chip manufacturers solve the compatibility problems between various frameworks, which brings great challenges to the versatility of AI chips. In the ecological aspect of AI products, the scale of the AI ​​chip is still being explored. At the same time, in the process of landing, the AI ​​chip is not isolated, but also requires software applications, solutions and support from service providers. The AI ​​chip is ultimately a product. Since it is a product, the delivery of the AI ​​chip may be the chip itself, but it is more likely to be a chip + application + service. If there is only a chip, but there are no chip-based algorithms and applications, then it cannot be called a complete product. For example, for smart speakers, as a solution provider, AI chips are only a small part of their products, while other parts such as AI chip-based applications, cloud-based intelligent voice interactive services, content services and services Support, etc. This is the more important part of landing. In the process of landing the AI ​​chip, the customer needs a complete solution. If only the AI ​​chip is provided, it is necessary to find a corresponding partner in various aspects such as application and cloud service, so that it is possible to provide a complete customer. solution.
In short, AIOT's human-computer interaction is a huge market, and this has brought huge demand for AI chips. However, in the process of landing the AI ​​chip, it faces many challenges such as research and development, product positioning and commercialization. In terms of R&D, it is necessary to perform deep iterative optimization on the actual AI algorithm to meet product requirements and maintain the flexibility of the architecture. In terms of product positioning, in view of the reality of IoT device fragmentation, it is necessary to first consider the application scenario and scope of application. This again pushes down the functional and performance requirements of the AI ​​chip; in terms of commercialization path, the end customer often needs the overall solution instead of the chip itself, so how to build a complete AI solution is something that every AI chip player must consider. Things.
Editor in charge: Zhang Zequn
Pet Carriers And Travel Products
-Comfortable & Breathable
-Better Care to Your Pets
-Safe and Fashion Design
-Enjoy Trips with Your Pets
EASY CARRIER AND EASY CLEAN – Wear this airline approved pet carrier over the shoulder, secure it in the car with seatbelt, place atop your luggage or carry it with the top handle. Versatile for your cars or all types of travel,and the cozy fleece pet bed can be completely removed and machine-washed.
Customer Experience Comes First: We always stand behind our customers and try best to meet your needs. We have made several upgraded improvements such as thickening shoulder pad sponge, using upgraded breathable mesh and polyester fiber material, which makes the carrier more durable and comfortable.
Enjoy Trips with Your Pet: Our pet carrier is lightweight and portable, you can use it to carry small dogs or cats for biking, hiking, camping, out walking or meeting friends. It allows owners and their lovely pets to have more adventures together, what's more, it's also a great photo prop to take funny pictures for daily sharing.
We are a Chinese pet products manufacturer, our company produces and sells various cat accessories and dog accessories, including Pet Toys , Pet Apparel And Accessories, Pet Collar And Leashes , Pet Beds And Accessories, Pet Cleaning And Grooming Products, Pet Bowls And Feeders , Pet Houses , and Pet Carriers And Travel Products, which are accepted OEM&ODM. Please don`t hesitate to contact me if you need anything.
Pet Carriers And Travel Products,Cat Travel Backpack,Pet Carriers For Cats,Pet Travel Bag
Jilin LOYO Pet Products Co.,Ltd. , https://www.jilinloyopets.com