Bitcoin plummeted by 10% in the first 24 hours and fell below 9000 US dollars;
In the past two days, for Bitcoin miners, it was a very difficult two days. In a short period of time, Bitcoin plunged 10% and once again fell below the 9,000 mark.
Not only Bitcoin, but the second largest electronic currency, Ether, tumbled 7%, and the third-largest electronic currency, Ripple, fell 10%.
There are certainly reasons for such a bad market. According to Bloomberg News reported on the 11th, South Korea raided Upbit, one of the world’s largest electronic currency exchanges, because of its alleged transfer of client funds to company accounts.
The former largest Bitcoin exchange in the world (bankruptcy) was preparing to sell its remaining tens of thousands of bitcoins, and Bitcoin collapsed.
It must be said that electronic money is indeed a financial tool to test psychological quality. Do you not know if bitcoin can return to the peak of 20,000 US dollars in the future?
2. Bitcoin suffered a three-jumping attack More opportunities in AI?
Securities Times reporter Wu Jiaming
In recent days, the three consecutive heavy blows have caused bitcoin prices to fall sharply again. At this time, some investors seem to be looking to another area: artificial intelligence (AI).
According to overseas media reports, the recent Bitcoin market can be described as bad news. The first crackdown came from Mt. Gox, who was once the world’s largest Bitcoin exchange but is currently bankrupt. Ms. Gox’s large-scale shipments were managed by the company, according to Mt. Gox’s The Cold Wallet monitoring website shows that about 138,000 Bitcoin remains in the Mt.Gox Wallet. The price of Bitcoin fell.
The second strike came from Nvidia. Nvidia's latest quarterly financial report for the first quarter of fiscal year 2018 showed that revenue related to mining operations reached US$289 million, but Colette Kress, Chief Financial Officer of Nvidia, said that the company expects revenue related to this business to fall by 65 in the second quarter. % To a certain extent, it also triggered the selling pressure of digital currency.
The third strike came from South Korea. According to an official from South Korea's Seoul Prosecutor's Office, South Korean prosecutors raided the office of Upbit, one of the world’s largest digital exchanges, on Thursday and Friday. It is reported that Upbit has allegedly transferred customers. Account funds go to company accounts. Upbit is in charge of Bitcoin worth about 1.6 billion U.S. dollars. It is Korea's largest and fourth largest Bitcoin exchange in the world.
The price of Bitcoin still fluctuates violently, which also brings uncertainty. At this time, some investors began to see other opportunities.
Benjamin Lau, chief investment officer of Apriem Advisors, said that Nvidia also has a bigger opportunity than mining. 'For artificial intelligence and data centers, the demand for Nvidia chips is strong, and this should drive the company's business development in the near future. " Benjamin Lau proposed that Nvidia's commercial chips for the AI sector could include healthcare, transportation and computer data centers. This is the future.
Coincidentally, in an interview with Bloomberg, Bitland’s co-founder, Wu Ji Han, revealed that the company’s Bitcoin mining business in the United States has a huge expansion plan. He also commented on the company’s position in artificial intelligence, calling it Bit Continental. Also planning to set foot in this field.
Today, AI has infiltrated into various fields. Recently, Total formally announced an agreement with Google Cloud, which will jointly develop artificial intelligence technology and provide a new intelligent solution for oil natural exploration and development. This incident immediately in the global oil industry Concerned. Since Alpha Dog can surpass the Go World Championship, will Total and Google create 'Super Smart Oil Man' beyond the average oil person, will the oil industry usher in a subversion?
In fact, the accelerated rise of artificial intelligence in recent years is having a major impact on the global chip industry. Nvidia has changed from a graphics card supplier to an artificial intelligence server provider. Artificial intelligence including Google, Facebook, Microsoft and other technology giants The research leader is already using the chip products provided by Nvidia specifically for research in this field. Market analysts predict that the global artificial intelligence market will be worth about 500 billion U.S. dollars over the next 10 years.
3. The potential of the future of the blockchain application is longer.
Bitcoin, Ethereum and other virtual currencies have become more and more popular. Blockchain technology has attracted more and more attention. Many industries have begun to think about how to use blockchain technology to create more value. , Even a new business. Among them, the development of supply chain finance may be the most important trend affecting Taiwan's industry.
In the traditional financial system, all records of transactions and financial transactions are stored in the central database of financial institutions. Therefore, financial institutions must invest considerable human and material resources to maintain their databases to ensure that they can operate stably and that the data is safe and secure. In terms of the financial system, related investments have also increased the cost of financial transactions.
In view of this, the concept of distributed ledger emerged. Under the distributed ledger structure, all parties involved in the transaction have a part of the entire ledger, so no central database is needed to keep the entire transaction ledger. However, due to the parties to the transaction All the books you have are only part of the full book. How to ensure that the other books are correct and that the data has not been tampered with will become a big problem. The blockchain is an effective solution to this problem.
There are various blockchain applications, even more powerful with Industry 4.0
Further deduction, blockchain technology can not only be used in the financial industry, any data using distributed architecture, can use blockchain technology to verify its authenticity. More importantly, blockchain technology is basically open source, Just on Github, there are 26,000 blockchain-based development projects. Well-known international companies such as IBM and SAP also support the open source blockchain Hyperledger Fabric, which can be used to implement Commercial applications.
As a matter of fact, the topic of industry 4.0 that has been prolonged for many years will inevitably make use of blockchain technology. Industry 4.0 has many facets, except for the concept of digitalization of processes, digital twins, and IIoT. Besides the terminology, at the later stage of the development of Industry 4.0, upstream and downstream information systems (such as ERP, MES, etc.) in the supply chain must also be connected in series so that manufacturers can achieve flexible production and reduce stock preparation.
According to the author's understanding, some large-scale accounting firms have begun to take precautions for the changes brought about by Industry 4.0, and have cooperated with relevant industrial equipment/system manufacturers. Because, in the highly digitalized manufacturing, raw materials, semi-finished products, and product inventory data The change will be in seconds. How to respond to these changes in real time in the accounting process and in the financial statements is also a big challenge.
On the other hand, manufacturing customers must be very sensitive to the above figures, especially large-scale electronics manufacturing, because the gross profit margin is generally low. Whether it is raw materials, semi-finished products or finished product inventory, everything is a cost item, and it is often the profit and loss of the company. The key point is that although securities regulators may not require listed companies to shorten the disclosure cycle of earnings-related information, real-time financial data is still an important reference for corporate decision-making.
For a large-scale manufacturing industry such as Taiwan's Electronics Group, it is necessary to use a centralized architecture to handle the data of thousands of upstream and downstream manufacturers. IT-related software and hardware costs must be very striking, and the system may also be It will become large and complex, and it will bring challenges to management and maintenance. Importing blockchain technology and adopting distributed data architecture will be an attractive solution for these large-scale manufacturers.
In addition to applications related to Invoicing management, blockchains have a number of other application possibilities. For example, in the context of industrial IoT, sensor networks constantly send out various data, but in the event of transmission of these data, After tampering, the operation of the production line may become chaotic and even jeopardize the safety of the factory. Blockchain technology can be used to prevent data from being tampered with and to further protect the information security of the Industrial Internet of Things.
Corporate finance will flip due to blockchain
In addition to verifying the authenticity of distributed data and ensuring that data is not tampered with, there is an important application of blockchain, which is to issue tokens. And this application is precisely what caused the virtual currency market to flood with a lot of bubbles. The reason - almost anyone can issue tokens to finance, use virtual currencies in exchange for dollars, renminbi, NT and other legal currencies with government endorsements.
Regardless of the various chaos and bubbles that exist in the current virtual currency market, tokens are actually an application that has deep potential for corporate finance applications. In the case of manufacturing, products cannot be created out of thin air. One manufacturer wants to produce certain items. For products, it is necessary to purchase raw materials or components from upstream suppliers, and pay the purchase price at a predetermined period.
In the current situation, even though the supplier has already shipped the goods to the customer, what they have got is a check that can be cashed in a few months. Should the supplier suddenly need cash flow, it can only find the bank discount, and There is also an interest to be paid to the bank. Moreover, check discounting is subject to many restrictions. For example, a check with a long ticket period is usually not accepted by the bank. This also allows the underground financial industry to have room to survive. In fact, many large scale Manufacturing industry uses the time difference between accounts receivable and accounts payable to create huge profits, but the capital risk is borne by the supplier.
On the other hand, if the customer abruptly places large orders, the supplier must urgently expand production. Unless the customer is willing to pay the purchase price in advance, many small and medium-sized suppliers may not be able to do so because banks usually have strict requirements for SME financing. Restrictions: SMEs can often get very low credit lines and can't even get credit lines. At this time, SME owners often only have a chance to take a loan and use their own real estate as loan collateral. Many are attached to large-scale loans. The SMEs under the manufacturing industry, although their businesses seem to be stable, have found that their cash flow is not working properly and they cannot adjust their positions with normal financial institutions. The main reason is here.
To some extent, this is also one of the main needs of the real estate market in Taiwan. If small and medium-sized business owners have spare cash, investing in immovable property is an option that can be attacked and retired. It is both an opportunity to earn a bid-ask spread, and it can also be used by companies. Working capital is used as a collateral for loans. Real estate is one of the bank's most willing to accept loan collateral, because its price fluctuation is usually lower than other securities, and the real estate appraisal mechanism is quite mature.
If large-scale manufacturers use blockchain technology to issue their own tokens and circulate within their own supply chain system, the entire supply chain's capital operation schedule will present another scenario. In fact, the generation of blockchain technology distribution Currency is not only a substitute for traditional currency, but also can be equipped with Smart Contract function. The issuer can not only customize the contract content with the recipient. After the contract conditions are satisfied, subsequent trading actions can also be performed automatically. There is absolutely no need for human intervention. This feature can significantly reduce transaction costs for buyers and sellers.
In other words, tokens with smart contracts are a bit similar to letters of credit (LCs) used in international trade, but there is no bank intermediary guarantee. And if these tokens have limited openness to the secondary market, Manufacturers' supply chain members can buy and sell members in the same supply chain system (this creates a bond-like market, but the participants are limited to a certain group of specific corporate entities), or they can use it as financing collateral. It will cause great disruption to the existing corporate financial market.
Moreover, according to the current financial regulations, it is difficult for the government to intervene in the regulation because the entire system is generated by the transactions between private companies. The “currency” or “credential” to be used by buyers and sellers as the medium of transaction is entirely two. Creating contractual freedoms. It is not uncommon for companies to borrow money and turn around each other. What assets should be used as collateral and even unsecured loans. If both parties are not listed companies, they are basically the boss.
The golden opportunity of traditional enterprise gold
It may be easy for readers to understand why the traditional financial industry is so cautious about Fintech innovation based on blockchain technology because in this new game rules, financial institutions are not only Marginalization may even be excluded.
However, for the traditional financial industry, the situation may not be so pessimistic. As mentioned above, due to various factors such as credit costs and risks, banks are basically hard to make credit loans to SMEs and must have collateral. If the large-scale manufacturing industry issues tokens for its own supply chain companies, and opens its algorithm to banks, banks can easily grasp the transaction records of small and medium-sized enterprises with large-scale manufacturing industries in the system, significantly reducing credit costs. The supervisory authorities are willing to loosen further regulations and treat these tokens as quasi-products of certain types of securities. Of course, these tokens can also be used to finance bank mortgages.
Regardless of whether the former is the latter or the latter, for the existing financial institutions, the emergence of tokens helps to make the pie of SME financial services even bigger.
In fact, if financial regulators are to intervene in large-scale manufacturing to issue tokens for their own supply chain system, conferring certain legal status on these tokens is an inevitable result. The legal status of tokens must be clearly defined before the government can proceed. To control. If the token is not legally anything, government regulations also go unnamed.
Long-term plan for mining machine industry
Since the exchange rate of Bitcoin against the US dollar soared in 2017 and officially detonated the long-awaited boom in the mining industry, the opportunity for mining machines is a hot topic within the electronics industry chain in Taiwan. There are also several IC design companies in China that have launched mining operations. Dedicated ASICs burst into flames. However, if you understand Bitcoin, the rules behind the game, you can assert that the current business model of mining for cash will not be a long-lasting business.
In terms of Bitcoin, its total number of issues is 21 million, and there will be no new bitcoin after the release. This is the “big limit” of the Bitcoin mining machine industry. It is estimated that as of mid-2017, bits The circulation of coins has reached 16.38 million, and as Bitcoin's remaining issuance quota is getting lower and lower, it will take longer for mining machines to dig into a bitcoin, resulting in a return on mining investment. Decreasing with time. It is generally believed that the remaining 4 million bitcoins will be issued before 2040.
As for Ethereum, although its total amount is not as bitcoin as Bitcoin, Vitalik Buterin, the co-founder of Ethereum, proposed to adjust the upper limit of the circulation of Ethereum to 140 million, which shows the mining of Ethereum. In the end, the machine may still encounter problems similar to Bitcoin.
On the other hand, the Ethereum community has always been opposed to mass production of the Ether coins by mining machines. Because the over-concentration of the cryptocurrency owners will cause many drawbacks to the cryptocurrencies themselves, professional miners, In particular, professional miners using ASIC excavators will accelerate the concentration of cryptocurrency ownership. Currently, TEC-based mining machines are used for excavation of coal coins. The mining efficiency is lower than that of ASICs, but the benefits are Adapt to changes in the mining algorithm.
A few days ago, China's IC design company Bitland announced that it will launch an Ethernet mining ASIC dedicated to ether, and a few days later, the members of the Ethereum Foundation (ETH) immediately proposed to modify the mining algorithm of the Ethernet coin. Obviously, it is aimed at ASICs such as Bitland. The developer's behavior comes. The outcome of this battle is very easy to foresee - the development cycle of ASICs is at least three to six months, and the cost of photomasks is often millions of dollars, but if there is consensus in the community, It may take only a few weeks to modify the algorithm, and the cost is close to zero. Hardware and software "resilience" always suffer, and even FPGAs with the most flexible circuit architecture have some limitations.
In general, the current business model of mining machines is not sustainable. However, if all industries and industries begin to develop their own vertical applications with blockchain technology, the situation will be very different. The mining behavior is essentially a solution. For example, to deal with blockchain data, therefore, with the wide application of blockchain technology, all industries and industries will have relevant needs. This is the long-term business of the mining industry. As mentioned in this article, In the case of supply chain finance, for example, hundreds of thousands of electronic related companies all over the world use blockchain technology to deal with supply chain financial problems. With the demand of banks and other traditional financial institutions, mining machines will become one of the corporate IT markets. The door is not a small business, but this market may be relatively fragmented and may not be suitable for programmable solutions such as ASIC, FPGA or GPU, and may be commercially feasible.
4. AI intelligence is omnipresent Smart terminals introduce large-scale edge operations;
After cloud computing matures, large-scale, extensive operations will remain in the cloud, and small-scale, locally-accurate features and demanding operations will shift to the edge or fog. In 2019, the global fog computing market will be approximately US$3.7 billion. In 2022, it will grow further to about 18.2 billion U.S. dollars. The concept of distributed architecture rises, and edge computing will become the focus of development in the coming years.
AI artificial intelligence computing technology will gradually develop from the cloud to the edge in the next few years. The application field is all-encompassing. The demand for intelligent functions of various portable devices and terminal devices is high, but the huge amount of Internet of Things will lead to data. Congestion can not all be handled through the cloud, so it must rely on edge operations, giving the terminal more computing power.
Edge operations can be roughly classified into Mobile Edge Computing and Fog Computing. AI-specific hardware/accelerators have mushroomed recently, and chips specifically designed for neural network arithmetic have emerged, including AI-Optimized. Processor, Deep Learning Processor, AI Accelerator, Neural Processing Unit (NPU), etc. The AI edge computation needs to simplify the neural network model, and at the same time, it strengthens the computational efficiency, and allows the models trained by general enterprises to be compressed to be used on mobile devices.
Distributed architecture concept The rise of edge computing
The concept of Edge Computing arises from the artificial intelligence (AI) boom. After the cloud computing in the old days matures, large-scale and extensive operations will remain in the cloud. Small-scale, localized features and precise requirements The calculations will be moved to the edge or fog. According to the Open Fog Consortium study, the global fog computing market will be approximately $3.7 billion in 2019, and will further grow to approximately $18.2 billion in 2022. The main application areas are Utilities, Transportation, Healthcare, Industrial, Agriculture, etc.
In addition, due to the development of the Internet of Things, the number of network-connected terminal nodes will show an explosive growth in the coming years, and a huge amount of Internet of Things will soon take shape. Wang Bingfeng, Deputy Director of the Center for Applied Materials Systems of CCDC (Figure 1) believes that Because data data will be generated in large quantities, existing transmission architectures cannot cope, congestion must be faced, and there is insufficient bandwidth; in order to handle large amounts of data, operators must build more cloud-based equipment, which keeps the operating costs of enterprises higher. With the rise of the concept of distributed architecture, edge computing has become the focus of development in the coming years.
Figure 1 Wang Bingfeng, deputy director of the Center for Applied Materials Systems of the Institute of Information Technology, believes that the data will be generated in large quantities and that the existing transmission architecture cannot afford it. Therefore, the concept of distributed architecture arises, and edge operations will become the focus of development in the coming years.
The cloud computing and edge operations have significant differences in technical architecture and many features. As shown in Table 1, Wang Bingfeng explained that edge computing will use a distributed architecture and operate on numerous fog nodes. The three major technical characteristics of operations include: The location of the operation, from the center down to the edge; Also due to the location of the operation on the terminal, so the network latency rate is low, can support applications of 10 milliseconds; For network bandwidth, infrastructure The requirement is small and can provide some services in the absence of links in the cloud network.
Edge operations will be deployed more flexibly when they are in operation. The cost of the edge node of the Internet of Things is relatively small and can be configured in a large number of points; the edge nodes formed do not require too complex deployment planning. Relatively flexible visual context adjustments. In the context of application, Wang Bingfeng explained that the development of the Internet of Things has facilitated a large number of applications for advanced queries and analysis of a large number of sensor data streams; the second is real-time event detection; third The Networked Control System (NCS) is the industrial automation system; the fourth is the Mobile Crowdsensing (MCS).
Edge and cloud computing complement each other without conflict
The combination of AI and IoT will drive the development of more intelligent systems. Existing cloud architectures cannot meet the needs of off-line processing, data privacy, and real-time response of the Internet of Things, and lead into the edge computing architecture to solve the problems faced by the cloud architecture and increase the flexibility of use. Solving the need for sharing and reusing resources between different systems. Wang Bingfeng explained that Edge Intelligence via Ambient Computing is defined as a device that is far from the cloud and close to the consumer's device. It must have certain computing and intelligence. Generates the ability to process nearby sensed data.
The development challenges of edge computing include: Scalability, Complex Inter-networking, Dynamics and Adaptation, Diverisity and Heterogeneity, etc. The scale of the Internet of Things will continue to expand, and terminal nodes will be physically connected with different conditions and types. Devices with wireless connectivity and mobility need to allocate edge resources in real time and re-embed IoT applications; edge devices also need to be Seamless interface with interoperability.
The emergence of edge computing does not replace cloud computing. Instead, it will develop in a complementary direction. With the immediate response of the terminal, the burden of cloud computing can be reduced. However, in terms of the overall IoT context, the significance of edge operations lies in In response to more heterogeneous situations, especially with the increase in the types of IoT devices, more sub-systems will be derived, and whether these sub-systems can effectively perform collaborative operations will affect the quality of intelligent services.
Deep learning drives AI growth
The wave of development of artificial intelligence AI is based on the development of the past, and it is the most representative of Deep Learning. According to Yin Siezhi, chief artificial intelligence scientist at DeepBelief.ai (Figure 2), the success of deep learning comes from deeper learning. Understanding the operating mechanism of the human brain, the core of which is characterization learning, where machine vision is to establish the mapping relationship between human vision and machine-readable pixels. Humans can easily understand the deep rules of the appearance of images, so most people can In the absence of data, make correct judgments and identify many computers that need long-term training.
Fig. 2 Yin Xiangzhi, chief artificial intelligence scientist of DeepBelief.ai, stated that the success of deep learning comes from a deeper understanding of the operating mechanism of the human brain. Its core is the representational learning.
So deep learning means that we begin to teach machines to deal with this world in complex ways by creating complex models and establishing rules for mutual assistance. Because the world is complex, if we try to use simplified methods and models, we are destined to Failure. Yin Xiangzhi explained that Convolutional Neural Network (CNN), Recursive Neural Network (RNN), Reinforcement Learning (Reinforcement Learning) are the three most commonly used deep learning techniques recently.
Convolutional neural networks are used to identify images. The famous Imagenet computer vision contest has become a convolutional neural network model design competition in the later period. Almost all participating teams have designed their engines using convolutional neural networks. In the application of edge computing, Face recognition is expected to be the mainstream of edge computing. There are more ways to understand customers. More and more applications have been introduced in China. Face recognition has been introduced, including train station face-to-face, mobile payment using face authentication, and universities. The dormitory's access control system also uses face recognition, and "brushing the face" will become one of the most popular image recognition applications.
Chinese recognition is also suitable for convolutional neural networks. In fact, Chinese is an image. The biggest difference from other languages is that it is difficult for Chinese to disassemble into simple symbols. In terms of language recognition, recurrent neural networks are currently the most suitable In 2016, Google Translate was completely replaced by the RNN model, and effectively improved the accuracy of translation. The correct rate of Microsoft's speech recognition technology also officially exceeded the speedometer for human professionals. In addition to Chinese, the recognition rate of other language computers worldwide has exceeded Humanity.
In reinforcement learning, Generative Adversarial Networks (GAN) is a fairly popular technology recently. How to determine that the computer has learned all the characteristics of the thing, GAN uses two sets of deep learning to determine whether it is a real picture or a fake picture, and the other It is responsible for producing false images that make it harder to judge whether it is true or false. When the recognition model is indistinguishable from true and false, it means that you have fully grasped the deep features of the thing. By using this feature, you can create real images that are real and use the GAN engine to capture any image. Replace it with another protagonist.
The use of natural language to compile knowledge maps, and knowledge maps to reason through natural language, is expected to be the next wave of growth for artificial intelligence. Yin Xiangzhi said that the White House annual economic outlook forecast report, hourly pay below $ 20 There will be a 83% chance that the work will be replaced by AI, and the primary data analysts will soon be replaced. However, the development of this wave of AI is not the disappearance of work, but in the liberation of brainpower, we do not need to devote much time and energy to Perform repetitive, low-value work, and do more high-level, creative work.
Optimize the architecture to enhance edge performance
Judging from the current application of technology and development, AI is not entirely applied to completely new devices. Instead, it first introduces existing equipment, including: VR/MR, robots, drones, self-driving, IoT nodes, smart homes. , Medical, mobile terminals, servers, transportation and logistics, etc. Arm's senior regional marketing manager Cai Wunan (Figure 3) said that the most common AI and edge computing terminals are smart phones, applications such as speech recognition, text prediction (Predictive Text), Face Tracking Camera, Digital Assistant, Augmented Reality, Fingerprint Identity, etc.
Fig. 3 Cai Wunan, Arm's senior regional marketing manager, said that in the application of deep learning, using the optimized neural network link library to run similar functions can improve performance by five times; the speed of the software can be increased by fifteen times.
Arm is a leading manufacturer of embedded CPU architecture. He did not actively invest in the development of AI several years ago. However, after AI showed a comprehensive development trend, the company also optimized hardware and software for AI applications, including machine learning. Processor (Machine Learning Processor), Object Detection Processor, Neural Network Software Link Library, etc.
Based on network bandwidth, power consumption, cost, real-time response, reliability, security and privacy and other requirements, edge computing will enter a stage of rapid development. Cai Wunan pointed out that in the application of deep learning, the use of optimized neural network link library operation With similar features, it can improve performance by five times; the software can run 15 times faster than before.
Cloud and Edge Software and Hardware Architecture Integration
The technical architecture of AI can be roughly divided into three parts: Training, Deep Neural Network model, and Inferencing. NVIDIA technical marketing manager Su Jiaxing (Figure 4) stated that AI's The overall development model has become more and more complex and huge. In the image recognition section, 2016 Google’s Inception-v4 is 350 times more complex than 2012's AlexNet; Baidu’s DeepSpeech 3 voice recognition engine was published in 2017, compared with 2014. The first generation is 30 times more complex; the translation engine 2017 MoE is also 10 times more complicated than the 2015 OpenNMT.
Figure 4: Su Jiaxing, NVIDIA's technical marketing manager, said that in recent years, the AI neural network model has become more and more complex and huge, and the cost of producing high-accuracy AI is getting higher and higher.
In the face of complex deep learning networks, to produce high-accuracy AI costs are getting higher and higher, including power consumption, computing power, network bandwidth and other resources. Su Jiaxing explained that NVIDIA's new inference accelerator also tries to simplify the architecture by merging. Repeated operation flow to reduce unnecessary calculations, simplify the hardware burden, optimize the hardware, from the 140 images per second to 5,700 images in the image recognition section, 40 times faster; the translation sentence also improves from 4 sentences per second To 550 sentences, increase 140 times.
In addition, the smart city is also a situation where the cloud and the edge are combined. After many terminal data are collected by the camera and the sensing node, they are quickly processed at the terminal and sent to the cloud for data integration. For the security of the city, crimes The prevention and treatment of disasters have greatly helped; Su Jiaxing believes that the future application of such large-scale integration of cloud and edge computing, especially in the integration of cloud hardware and software with edge software and hardware, and AI computing, will increase. The higher the price, the more emerging market is full of business opportunities.
AI Accelerator Development Reaccelerates
The goal of most of the development of science and technology is low power consumption, low cost, high efficiency, and the development of AI is no exception. It is also the reason for the rise of edge computing. Taking the increasingly common image identification technology as an example, Feng Chia According to Chen Guanhong, an associate professor in the Department of Electronic Engineering at University of Hong Kong (Figure 5), about 90% of the convolutional neural network architecture is concentrated in the Convolution Layer. Therefore, reducing the complexity of the convolutional layer can effectively reduce the computational complexity of the convolutional layer. Inferring the computational burden of the hardware, besides the above-mentioned offline operation, the data reuse is also one of the important principles.
Fig. 5 Chen Guanhong, an associate professor of electronic engineering at Feng Chia University, shows that about 90% of convolutional neural networks are concentrated in the convolutional layer, which reduces the computational complexity of the convolutional layer and can effectively reduce the hardware computational burden.
In addition, the role of the activation functions in the neural network is to add some nonlinear factors to the neural network, so that the neural network can solve complex problems, make the results closer to the habits of human decision making, and it has nonlinearity and differentiability. , monotonicity and range of output values, usually ranging from -1 to +1, and the result will be close to the middle, the higher the number of iterations in the neural network operation, the more accurate the result will be. Acceleration from the network and hardware performance From a perspective, there should be more and more dedicated AI accelerators or AI chips coming out in the future.
MIT's research team has published a chip "Eyeriss" that specializes in deep learning. It can directly execute algorithms such as face recognition on mobile devices and can process data offline. Explain that Eyeriss chip has 168 cores built specifically for deploying neural networks. Its performance is 10 times that of general mobile GPUs. Because of its high performance, it can directly perform artificial intelligence on mobile devices without processing data through the network. Algorithm. Ability to recognize face and language. Can be applied to smart phones, wearable devices, robots, autonomous vehicles and other Internet of Things applications. New Electronics
5. Yun Zhisheng CEO Huang Wei: The impetuous era of AI is over. There will be a watershed in 2018.
Editor's note: The guest of this session was Yun Zhisheng's founder, CEO Huang Wei. How did he build the cloud pyramid's R&D model? After six years of entrepreneurship, how did he see the status and future of the industry? How did he hear him together? Say!
There are many things happening in six years, and time keeps washing every industry.
Yun Zhisheng, founded in 2012, is more like a Long March contingent that is advancing toward its destination. It has a clear mission, invests its troops, and steadily progresses according to the established operational plan.
This is the origin of the founder Huang Wei’s self-confidence. According to statistics, there are more than 20,000 partners in the cloud, more than 200 million users, and 3.3 million calls on the cloud platform, covering more than 647 cities.
Huang Wei graduated from the University of Science and Technology of China and received his PhD. After graduating, he worked as a senior researcher at the Motorola China Research Center. He developed the world's first mobile phone voiceprint authentication system. Later he became the core executive of Shanda Innovation Institute and created a voice. Branch; followed by the establishment of Yun Zhisheng in 2012.
He participated in the National Institute of Standards and Technology (NIST) speaker recognition evaluation (SRE) project for three consecutive years and won the first place in the main task. He is the only Chinese who has been doing Keynote Speaker for two consecutive years in the NIST evaluation.
'Pyramid' style R&D plan
In his opinion, Cloud's 'Pyramid'-style R&D plan has ensured that they have been in the industry for the past six years and that it will continue to do so for the next six years because their 'infrastructure' is perfect.
So, what is the 'pyramid' type of R&D? The bottom layer of the pyramid is the DeepFlow cluster. This isomerized hardware server cluster can provide intensive computation and storage capabilities upwards to ensure that the R&D team has sufficient computing support, Huang Wei said. In 2018, it plans to expand to more than 1000 GPUs.
The middle layer is the Atlas supercomputer platform, which is a distributed machine learning parallel computing platform. The internal collaborative and shared AI underlying research and development technology results can be migrated and reused in various fields.
“We made deep learning in 2012. This is the big data direction that we have determined. We will be able to fully engage in deep learning. I can say with responsibility that at that time, 95% of university professors in China had never heard of deep learning. ' Huang Wei told Netease intelligence.
The top layer of the 'Pyramid' is the application layer technology, such as the output of application layer technologies such as ASR, TTS, NLU. Huang Wei said that it can be understood that Yun Zhisheng has established a core capability. This core capability is AI, and then AI uses a cloud-core product architecture to produce it and then apply it to different application scenarios.
When it comes to applications, 2018 is called the first year of artificial intelligence, landing will become crucial. Huang Wei’s focus is on areas such as IoT and medical care. 'We need to release the accumulated potential energy. There is no doubt that this year. We will ship large-scale shipments in smart homes and robots'.
Deducing AI chip R&D from application scenarios is the best path for AI startups
At present, artificial intelligence chips are popular, whether they are traditional chip giants such as NVIDIA, Intel, Qualcomm, or AI chips Apple AI1, Haisi Kirin 970, Qualcomm Xiaolong 845, Samsung Exynos 9810, MediaTek Pali P60, and startups. The company's skyline journey chip, etc.
We gradually smell the taste of the Red Sea. The reason is that the chip is a very important computational carrier for artificial intelligence, as well as a sensory carrier. It carries a very important part of artificial intelligence and is essential.
However, Huang Wei believes that everyone does chip, but the chip actually includes several major categories. Everyone has overlap but each has its own market and force point. For example, Nvidia's GPU is used to support extremely high computational load. The question is the battlefield of the giants; there is also a cloud chip, which is mainly used to make some cloud-based cognitive decisions, such as Google’s TPU, etc., and there’s a cloud knowledge, what the horizons do. This type of chip is a terminal chip.
The most important thing to do the terminal AI chip is to define the product form and logic of the entire chip from the application scenario.
Huang Wei pointed out that in the form of chip products, Yun Zhisheng has been exploring for a long time. Starting from the market, Yun Zhisheng has been based on IVM (Universal Chip Solution) in smart home, smart speaker, and children's education robot market. The product form validates the rationality of the market, product, and user scenarios. The higher the voice, stability, integration and other aspects of the product, the higher the number of product types and modalities, the more the introduction of self-developed AI chips becomes A matter of course.
Huang Wei stated that as early as 2015 Yun Zhisheng had formed a chip team. In 2016, the company began to evaluate the market, products, technology and downstream partners of chips. In 2017, the definition of chip products was started, IP selection, and algorithm optimization. , Tool preparation, and detailed product definition and technical module evaluation. At present, the UniOne AI chip has been taped out and will be released soon.
In terms of the chip, we are at least three years ahead of time. The possible timing is just in time for the AI chip hot spot, but this is also a reward for our early investment, 'yellow' Wei told Netease intelligence.
The first AI chip for IoT human interaction scenarios
As for the positioning of Yunzhisheng chip, Huang Wei stated that under the background of human-computer interaction of IoT, AI algorithm puts forward higher requirements on the parallel computing capacity and memory bandwidth of device-side chips. The traditional chip architecture is in these two However, on the other hand, despite the fact that GPU-based algorithms can implement inference algorithms in terminals, the disadvantages of large power consumption and low cost performance cannot be ignored. However, unlike mobile phones, IoT devices are in ever-changing forms, and demand fragmentation is severe. It is very difficult to cross device morphology problems. Therefore, only from the IoT application scenario, designing a customized chip architecture can significantly reduce the power consumption and cost while improving the performance, while satisfying the AI calculation and cross-device form requirements. .
At the same time, whether it is a large product or a small product, with or without a screen, the common things related to human-computer interaction should be extracted and solidified. In this sense, the chip is the most appropriate way. Our chip can be said to be the first AI chip for the IoT human-computer interaction scene. "Huang Wei said that from a commercial point of view, all the chips on the market today are not designed for artificial intelligence. It is difficult to use the power of the algorithm.
In addition, the module assembly mode that everyone adopts has the advantage of being able to ship quickly, but because of the cost constraints, many times it is only possible to choose chips with poor computing power in the market, and after the patchwork, there will be low yields. And other issues.
Therefore, Yun Zhisheng’s approach is to make terminal chips from the perspective of an algorithm company. It does not need to have particularly high computing power to build a sufficient chip. They think that if they do more, the cost and power consumption All aspects are wasted. And their chip selection speeds up calculations. Acceleration of the algorithm means that you limit its capabilities, and accelerating the calculations is 'learning internally.'
Huang Wei emphasizes that the most important thing is the positioning of the chip rather than following the trend. What is the ecology of the future? Where is the customer? Because the chip is a high-input, high-risk industry.
The impetuous era of AI is over. There will be a 'watershed' in 2018.
Huang Wei spoke in a circle of friends: 'I am old and I often wake up in the middle of the night'. In the face of reporters, he admits that there are also reasons for entrepreneurship. As entrepreneurs, they need to face from investors, customers, employees, markets, etc. When it comes to responsibility, he can't wait. He can't stop.
The smart hardware industry has experienced, and during the impetuous period of the past two years, it has begun to return to rationality. Huang Wei believes that 2018 will be the 'watershed' of the industry. Everyone starts to pay attention to landing. Competent companies begin to release potential energy, and they have no first-mover advantage. For companies, if you don’t have anything subversive, it’s hard to get the attention of investors and the market any longer, because the time window is over, and in the mainstream of the artificial intelligence industry, there is a unicorn. Beast, so the resources will be tilted to the head company.
'Commonly speaking, in the past few years you can still rely on bragging for investment, but this year has not been so simple,' Huang Wei said.
Regarding the development of artificial intelligence industry and human-computer interaction, he said that human interaction mainly depends on the ears and mouth. Vision allows us to see more of the content of the presentation. Speech allows us to express our initiative. Both are indispensable. , It complements each other. Therefore, voice and images must be the main interaction method in the future. Keyboard input or mouse input is essentially a violation of human nature. The future of mobile phones is not necessarily our necessities.
According to the data of the research agency GfK, the retail volume of smart speakers in China was only 10,000 units in 2015, and increased to 60,000 units in 2016. In the first eight months of 2017, a total of over 100,000 units were sold, with the third quarter of 2017 With the introduction of new products, only in August 2017, the smart speaker market reached a year-on-year growth rate of 178%.
The interactive revolution has come.
Of course, the public still has doubts about artificial intelligence. The UPC’s recent auto-pilot driving deaths have caused a lot of panic. This has caused everyone's panic. The AI threat theory has once again been heatedly debated by everyone. In response, Huang Wei reminded that technology should be safe. First, between the advancement of science and technology and the safety of human life, there is no doubt that the latter is more important. Financial News