Concerning Mitchell's question on the 'status of the development of AI chips', Yang Lin said that they have developed a USB-connectable hardware that can be used for AI calculations. He believes that we now need to 'play' AI. Only in this way can we embrace development; Li Weixing believes that mobile platforms such as mobile phones have great opportunities for the development of AI technology. The output of mobile phone chips is even greater, the cost is lower, and the degree of integration is also higher.
In the following links, Yang Lin continued to reiterate his claim of 'playing AI together'. He believes that through AI hardware, we can develop a variety of useful products. He took his cooperation with HKUST News and Communications as an example. Shows how to integrate speech recognition into AI hardware. Yang Lin believes that from students to retirees can participate in the development of AI.
Li Weixing said that the use scenario of AI is very rich. With the development of mobile platforms, both mobile phones and automobiles will become a terminal of AI. Yang Lin added that this year's mainstream mobile platforms can provide universal computing power of 0.1T. This is not enough for large-scale AI calculations, so it should be applied to mobile phones on the cloud. However, the overall idle capacity of Chinese mobile phones is very large. It is about 800 units of the power of Shenwei and Taihu Lake. We can Through the network, use these calculations.
Mitchell asked, In the just-concluded conversation, everyone raised the future-oriented computing methods such as the combination of distributed computing, cloud computing and local computing. But what is the biggest uncertainty in this future?
Yang Lin said that the development of communications technology has made these calculations possible, but the lack of killer applications may be the biggest uncertainty. Li Weixin continues to reiterate his claim that he will play with AI. He believes that 'hops' is so small. The game may be an opportunity for AI development. (Zhong Kai)
The following is a record of the round table:
Moderator: Please have Li Weixing, Yang Lin, Tom Mitchell conduct roundtable forum.
Host (Tom Mitchell): We're going to discuss a very important topic on the round table forum. We will talk about AI and hardware, including chip technology, how it will change the nature of artificial intelligence. I'm very pleased to have invited Yang Lin today. It is a chief scientist of (GTOI) and a professor at Berkeley. The other is Li Weixing, who is Qualcomm. First of all, I asked this question. We are still in the early stages of AI hardware provision. I would like to ask two people. How do you see some of the corresponding hardware trends supporting AI in the coming years?
Yang Lin: Now everyone knows that artificial intelligence has now entered the stage of industrialization. From the R&D to the industrialization stage, several new platforms are needed, and a new approach is also needed. This time I’m mainly talking about that now The platform is now NVIDIA's GPU, including Google's TPU. Now there are also Apple, Qualcomm, or Huawei's mobile phones. We have now developed a chip that can make a very small chip outside the phone USB, very low power consumption , The performance is very good. I now hope that everyone will play with artificial intelligence. Once it is time to play, we can really achieve industrialization. We say that AI is the subject of this conference. How to create everything, that is, there are various forms of landing. Platform. Below I will answer other specific questions.
Li Weixing: Thank you very much for the host's question. This is a very good question. Actually, I intend to answer this question. I also intend to answer it in Chinese. Actually, from the advancement of human science and technology, we enter into products that individuals can use. In fact, this The road is a very important one, from the technology, from the laboratory, from the school to the end is a special IC, and finally the products that everyone can use.
This road is very important. This road has actually been discovered in the previous 20 years. On the mobile platform, the mobile computing platform brings a very good chance. The most familiar mobile computing platform is the mobile phone, from the simple From the function machine to the slightly more complicated function machine, to now everyone's hands-on smartphone, the speed of advancement of the smartphone is very fast. There are a lot of technical problems that must be solved to provide some functions from the earliest Voice, camera, video, internet. Every different demand will naturally bring engineers some challenges. Through such a large platform, very fast iterative improvements, all good technologies, good production processes Can play on the mobile platform. This is the first advantage.
The second advantage is that because of its large volume, we say that we are called economies of scale. Let the cost quickly reach a key that everyone can afford.
The third, very important, is a reasoning of the mobile platform. It is a highly integrated requirement. Why? Because the mobile phone is not as big as the oldest big brother, it cannot be so convenient. You must reduce it. The advantage is that the power consumption is reduced, not only the standby time, but also the use of time.
Through the addition of these three very important factors, the use of methods has also improved, from voice to the wireless Internet, and now everyone's use of APP has changed everyone's online habits and changed everyone's spending habits. AI is another very important way to use it. It talks about the needs of the entire CPU, GPU, DIC, or hardware specially designed for AI. This is brought about by a highly integrated, rapid iterative economies of scale. Therefore, Qualcomm believes that this kind of AI in the field of action, and even other different ways of using it, will bring economically important reasoning.
The host (Tom Mitchell): Thank you very much for the very interesting point. As you can see, just now both pointed out that there are some very big changes in the basic hardware of AI. So I asked such a question, do you think there is Which, what new applications today can't be done today, but because of these new hardware, it may come out in the future.
Yang Lin: AI just talked about it. Now, on the one hand, it is R&D. We China has already established that Baidu is engaged in smart cars, Tencent is engaged in smart medical care, Ali is engaged in a smart city, and HKUST is engaged in language translation. However, this is a national major project. Just pointed out what it was called and played artificial intelligence. I can give you an example along these four aspects. That is, we can all participate in the creation of artificial intelligence applications. Just listening to the host of the conference also said, GMIC has been changed from the global Internet, or the future trend is machine intelligence.
So all of us are equivalent to what we want to go further. Let me talk about smart medical. Smart medical treats everyone. But if you put it in a mirror that you wake up every morning with a simple and inexpensive thing, it will watch you The family, Every day your status changes. This is not a matter of privacy, nor is it to be uploaded. But it will feel like you are not awake today, or if you are having a high fever today, or what is it. Then advise you Some other ways. Say that AI is producing everything, it's actually equivalent to the eyes of all things. Because the first thing is to ask the sensor. This is the smart medical problem I raised, and of course there are other issues, including infants and the elderly. guardianship.
Second, smart cars, we are not saying that apart from smart driving, what else can be done outside of intelligent transportation? Maybe it means what to do if you don’t find a car with a car key, or if your key is lost. You can use it there. Pattern recognition, face recognition, this set of things. Someone locked your car, another is this is our family, if not, you can spread the criminals to your mobile phone and tell you that the car was stolen. These things are not in the scope of smart driving, but they are also related to cars. They are developed by you. No one is developing for you. You can play such a thing yourself.
There are smart cities. Smart cities are certainly a very large area. We will mention smart homes. Domestically, I think there are also sweeping machines. What is a sweeping machine? There are smart sweepers now. After charging, I went for a charge myself. But we know that today the sweeper has no eyes. It is equivalent to hiring an old woman to help you sweep the floor. The efficiency is not very high. You can make a smart one according to whether your home is a carpet or a floor. The device, which is specifically where the dirty scan, save electricity, and see where you can sweep away. Now there are 1500-2000 yuan smart lock, this thing we can do. Smoke alarm, these smart home things We can all do it. This is one of the biggest breakthroughs in smart cities.
The last example is to fly with HKUST, and now everyone is engaged in translation on the Internet, or it is a speech comprehension. After the recent Facebook problem was created, it was not their own. In order to do so, there may be some people who use technology to exploit the loopholes. Take personal privacy away. Actually put this thing in a small thing of your own. For example, in the USB plug-in that I just took, you can translate your own voice recognition and give it training. And this is very simple, because there are two kinds in speech recognition, one is called specific person identification, which is for your own; the other is called non-specific person identification, which is based on all the people in the world. The recognition difficulty of non-specific person It's very, very big. Indeed, this small chip can't do it. But for your family, I'm a Beijing accent, Beijingers are training Beijing, and Shanghaiers are training Shanghai. This is a very simple question to identify. However, the privacy issue does not exist.
So Tom said these things just now, saying that it is the following how we are doing in industrialization. I feel that the government research projects on the one hand; on the one hand, we high school students, college students' retired intellectuals, can all develop the labor of our daily life. Intelligent hardware.
Li Weixing: This question is also a meaningful question. Just now Professor Yang’s various ways of using it, I’m talking about, from the nearest, to a little further, to your home, smart home, to the city, to the entire large region, It was just mentioned by Tom. For example, medical treatment, whether it is treating infectious diseases, is a variety of different ways of using it.
I discuss this issue from another angle, that is to say, there are actually a lot of things now. For example, the technology on our mobile phones and cars, maybe a professor with a PHD student 20 or 30 years ago, The PHD papers made by the very difficult computing platforms. Why are these things suddenly universal now? Actually because the computing platform has this ability. I'll take it as my first job in Qualcomm, which is how to use voice. Digitally. Because we used to call it a bender. We had to compress it when we reached 2G. At that time, the papers were discovered 20 or 30 years ago. We found that there are various techniques for voice compression. These concepts are, but not A computing platform, many good ideas can not be realized. But now turn this direction, once the market provides an effective, low-cost, very efficient computing platform, all of a sudden you would not have used the way you actually would have it. come out.
I often talk to my colleagues around me. When several Intels were invented by Dake, when I decided to invent Intel, I never thought that Facebook was a requirement of its design. Until this platform became accessible to everyone, it was natural to have The wise person uses this thing according to his personal or work situation. Qualcomm believes that the various uses of AI must be very rich. This is why we think it will be called whether it is hardware or software. The important thing is the framework. To make it available, use a very convenient way to integrate, so that the industry, and even the use of cross-industry methods have a chance to do it. This time there must be a way for everyone to use their eyes will naturally come up Qualcomm believes that mobile must be a way of using it. At this time, the so-called training and reasoning on the cloud, the computer, the edge, that is, the place on the mobile phone side, and even the car, will be a balance. I personally look forward to it. There is more meaning to make our life more convenient, and even heuristic use methods will be introduced in the industry. Thank you.
Yang Lin: I added some of what Dr. Lee just said. I very much agree with what he said. This is how to play it. Everyone now has a mobile phone. It is to use a cell phone to play it. Why do you use a cell phone to play it? I say here One, we have just said that artificial intelligence is a computing power. For the past ten years, mobile phones have been ASICs. Why? Is the power required for communication is very large. But ten years ago we all changed to general computing power. Arm, or other general-purpose chip. When the universal chip is counted up, we have a universal computing power into every one of us. What is this calculation? According to my estimation, it is new this year. Model, 0.1T, what does this mean, that is, the 10th to the 11th power of the second, is the simplest, artificial multiplication of this kind of addition, addition. As you can see, this is a huge calculation, but artificial The intelligence required for intelligence will be much greater than this. This is why Dr. Lee said in the cloud that mobile phones are used. I also recently discussed this with other people. That is the problem of college education. How about college students? For students, it means that we also need to participate in training. How to participate, I would like to say one thing, that is, for three consecutive years, China is the leader of supercomputers. Our Taihu Lake, China, is 1.1 billion times per second. What is the concept? 10 seconds to the 17th power. This is equivalent to a Chinese Taihu Lake equal to 1 million mobile phones, which is quite large. However, China Mobile has 890 million users, so we are equivalent to 809 million Taihu Lake in China. At least half of the time, when we don't have time to sleep, nobody uses it, and there is no problem with electricity. Charging treasures are available to everyone. They can be used outside. Later, a mobile senior expert told me that After statistics, it takes only three hours per person per day. So when it comes to the use of one-eighth, seven-eighths of the time is idle. In each of us, I’m counting now, that’s the whole China Mobile’s hands quite There are eight hundred satellites in Taihu Lake, China, which are unused. So this is one of the most important platforms for us to play AI in the future. I am now for this ecological chain, for us now, as the first one with Qualcomm. Cell phone As a general platform, our dedicated accelerator is a landing platform. We are now using one of the most fashionable blockchain technologies. Blockchain technology can use all of our resources. There is no privacy issue. There is no one who engages in the application just to collect data. The person who is training is only an expert. He works in the daytime, develops applications for you at night, and does not use the company's high-end GPU or server, but uses his own mobile phone. So this soon, I believe that within two to three years, the popularity of AI will be popularized.
Li Weixing: I just added, or clarify, Professor Yang. Just mentioned the figures, from 2017 to 2021, is the next five years. In fact, the global sales volume of mobile phones is estimated to be 8.6 billion. Such a large amount, I have just The economies of scale mentioned and the speed of the whole iteration will extend a lot of technology forward.
The second point, I just had to talk about it in depth. Above the cloud, the so-called training and reasoning, everyone will think that it seems to be doing training on the terminal. In fact, Qualcomm does not think so. Qualcomm believes that there is also a need for training and reasoning on the terminal side. Because there is a lot of personal data, in fact Professor Tom mentioned just now, privacy is a very important issue. Because some important names are taken away, maybe it is not a big problem to pass on the Internet, but there are some very personal things. , must be on the side of the person, this time on your mobile phone side, through your own data to do training, to be able to properly use it. Of course, this also has the so-called low latency requirements. At the same time, it is also a communications company, so we are talking about 3G/4G. Actually, in 2020, everyone is now moving in the direction of 5G. But at the same time, the so-called data-down and down-delay are still not negligible. It is very important to do a lot of operations locally, whether it is the so-called local training or something. The third very important point is reliability. In the event of a power outage, I am not talking about a blackout, but a power outage on the Internet, or for other reasons. This time the demand at the local end is obvious.
The host (Tom Mitchell): The two describe a very beautiful future. Now I would like to ask a question. The two of you have just painted a picture of the future. Such a future picture is a distributed computer environment. We are now Already have a cloud server and terminal mobile phone, the future picture is a combination of the two, may also some USB is received inside the phone, is local. There are 5G can be very good with the cloud service.
One thing I would like to mention is that we may have a lot of thinking in this area. If we understand the entire infrastructure of the future communications, I would like to ask two questions. Which one is the largest one? In other words, what is the greatest uncertainty in achieving such a goal? Can this question be understood, and what is the greatest uncertainty? To achieve such a goal?
Li Weixing: I think there are actually some things here that are certain, including the communication network going forward, and then the computing power is basically Moslow getting smaller and smaller, representing higher and higher computing power. But if you ask me today, because we are in every one 2G to 3G, 3G to 4G will in fact ask a question, that is why we use 3G, why do we use 4G, why do we use 5G. Finally we will want to say, because killer applications or killers use 2G/ What did 3G/4G become? Like a smart phone that actually existed before Apple, but it seems that at one point in time it suddenly became that everyone could have it. If you ask me today, I think this is not known. Maybe it is all.
Yang Lin: When it comes to uncertainty, I am a more optimistic person, so I think it may be faster than the popularity of imaginary imaginary people now, and the final result will be better. Why? I joined last week A global software development conference, all for young people, is a jump. This small game is a young man. Our generation is a part of what I can't call workaholics, but I always feel that only work can Value. People think that play is also the biggest part of life. When I came down with them, I talked to them, that is, how to get them to participate in AI, that is, artificial intelligence. The first one is that we already have a mobile platform. , I also added a AI-specific accelerator to the mobile platform. How to play? The developer jumped and said, I'm going to add you this thing. I told him why he was involved and why I said that this is NP Hard. NP Hard is the same as Bitcoin. It is a hash code. I think why people are involved in bitcoin mining? First, Bitcoin needs a lot of power, so it needs Dedicated chip. The development of artificial intelligence is actually a NP Hard problem. It is learning and training. He wants to find this suitable thing.
Whoever does this job is to do it for the whole people, not a big company. A big company just does a very specific, it is also very important, but it takes a long time. You can imagine, smart cars, autopilot, Smart City, smart medical, this vision is very big, when can it come? Five years, ten years? I don't know, I'm not sure. Everybody, I personally think, these little smart homes like I just mentioned, smart home, You learn Chinese medicine, identify what Chinese herbal medicine is, these things are very, very convenient, and there is no big company to develop this thing for you. This is an interest that you need to play. So I personally think that all these people who play mini-games are I assure you that they must be playing in the same process as the AI app. Now it's an applet, an applet, and it's now a small AI. So the people I talk to are Alibaba, Tencent, Xiaomi, these people listened to them. They would create an ecological chain, spread all of these AI applications to high school students, make them accessible to everyone, and have no culture. To play.
The host (Tom Mitchell): Thank you very much, but we have reached the point of time. Here I say that our discussion has made me more aware of the fact that there are opportunities for us to face the same problems that are now abundant. Colorful, so this topic is very interesting. It can be said that the implementation of the hardware may also involve some privacy protection, it will also involve the delay and the interface and many other issues. Also hope that our audience will come together Our experts, thank them for their insights. Everyone applauded, Thank you.