Andy Li, Director of Business Development, Qualcomm: AI is not a "winner-takes-all" and Web3 is not a "flood beast"

Original author: Chen Xiaorui

原文来源：dethings

On April 18, Aethir launched the Qualcomm Snapdragon 865 chip for Aethir Edge products, powered by Qualcomm technology, during Token 2049 in Dubai.

Qualcomm is a semiconductor and communications technology company founded in 1985 and headquartered in San Diego, California, United States. The main business includes the R&D and sales of mobile communication chips, as well as the authorization of related patented technologies, which are widely used in a variety of wireless communication devices such as smart phones, tablet computers, automobiles, and Internet of Things devices. Aethir is a decentralized GPU cloud infrastructure provider founded in 2021 and headquartered in Singapore. According to public information, in July 2023, Aethir completed a $9 million Pre-A round of financing, which was valued at $150 million at that time.

Recently, DeThings interviewed Andy Li, director of business development at Qualcomm, and the following is the full text, which has been edited to a certain extent.

DeThings: Can you briefly introduce Qualcomm's technology and business?

Andy Li: Qualcomm has accumulated a lot of technology over the course of more than 30 years, and we are best known for our technologies in the field of mobile communications, from the earliest CDMA to UMTS and CDMA2000 for 3G, to 4G LTE, and now 5G. Qualcomm recently released a 5.5G evolution plan, which we call 5.5G, which is actually the evolution of 5G, and our lab has successfully implemented 6G technology, of course, the specific degree of commercialization depends on the market.

In addition to mobile communications, we also have technologies in computing such as CPUs and GPUs, such as Snapdragon CPUs and GPUs. Our GPUs strike a good balance between high performance and low power consumption. Qualcomm has had its own CDSP architecture for a long time, formerly known as CDSP (C stand for computing), which is a DSP that does computing. With the continuous development of AI algorithms such as computer vision, CNN, TensorFlow, etc., in recent years, Qualcomm has been adapting its CDSP architecture to the latest technologies. Recently, we also introduced Qualcomm NPUs, including matrix accelerators, tensor accelerators, and vector accelerators, with many optimizations for memory sharing bandwidth and access.

So Qualcomm's technology is not only mobile communication, but also technology such as computing. The first step in applying these technologies to the Internet of Things is to connect objects, which requires the use of Qualcomm's connectivity technologies, including not only cellular network connectivity, but also near-field communication technologies such as Bluetooth and WiFi. When thousands of objects are connected, the problem of data silos is solved. Because each object, although it may be just a small sensor, is constantly generating data. In the past, when not connected, the data of each device existed like an island, requiring manual data pulling and analysis. Once connected, this data forms a network, and at the network level, the AI model can automatically learn from this data and analyze it to extract value from it.

DeThings: From the PC era to the mobile Internet era, a new term "Internet of Things" (IoT) has emerged. What is your take on the term "Internet of Things"?

Andy Li: IoT is not a new term for Qualcomm. The Internet of Things has actually been around for a long time, and the Internet of Things, as the name suggests, is the connection of objects into a network, which is the first step. We have a very long history in this area. Everyone's perception of Qualcomm may be to do mobile communication-related businesses, such as the Snapdragon platform on the mobile phone, but in fact, Qualcomm's business is very extensive. Qualcomm sees itself as an enabling partner across a wide range of industries, leveraging our product and technology portfolio. The more well-known Qualcomm product is the Snapdragon brand, and recently released PC and XR (VR/AR) branded products.

Both Depins and the so-called Web3 are essentially the same as the traditional Internet of Things, i.e., connecting all the devices. The difference is that in the past, due to the weak computing power of each device, a powerful centralized computing platform was required to control. But as the computing power of each node increases, decentralization, or depins, becomes a possibility, because decentralization can only be achieved if each node is strong enough, otherwise a strong central brain is required for processing. When every detail has a certain amount of computing power, it is possible to build a flat network. That's where Qualcomm's opportunity lies, and we have a combination of technologies at the edge and on the device side that can give each node a powerful capability, plus connectivity technology to connect them, and that's where the so-called decentralized network can really be realized.

DeThings: After the advent of Chatgpt, everyone's attention was attracted to AI, what does Qualcomm think about computing in the AI era?

Andy Li: I think AI could really be the intelligent productivity that can change the production structure of the entire human society again, and thus the entire social structure of our society, since the second industrial revolution. However, the average person may have a generalized understanding of AI, which is artificial intelligence. So how to implement artificial intelligence into concrete life and make a concrete thing?

First, we may have seen general-purpose generative AI like ChatGPT. But ChatGPT represents a general-purpose AI, which is actually a general knowledge. To be liberal means that you need to know everything. As a result, the parameters of general-purpose AI like ChatGPT will grow in size, with models with more than 200 billion parameters already emerging.

But going back to the IoT or our industry end space, my point is, whether you really need a general-purpose generative AI like ChatGPT, that needs to be discussed further. Because in the IoT field or on industrial end devices, they are not used to do everything. When designing a device, we already think about the limited number of things it needs to do in the field it faces. Just as we don't ask the radio "what should my wife give for tomorrow's birthday", we only ask about the operation of the device. In this case, do we still need a general-purpose AI? I don't think it needs to, it needs a customized, vertical AI model for the domain in which it is located.

In doing so, we are further focused on the scope of AI. In terms of hardware requirements, it will also be reduced accordingly. Because if you want to run a model with tens of billions of parameters, you can only run it in the data center at the moment. On a phone, you may only be able to run a model with more than a billion parameters. For IoT devices, a small model with 7 billion or even 4 billion parameters is sufficient. Instead of stacking hardware to run large models, we need to focus on how to make AI models run more efficiently.

Qualcomm also has a related layout in this regard. In addition to continuously improving the capabilities of our chips so that devices that were previously unable to run large models can now run, we are also introducing Qualcomm's AI framework. It should be noted that Qualcomm does not make large models, but provides a complete set of tools. We launched the AI Hub last week at the Embedded Show in Nuremberg, Germany, which actually ported hundreds of AI models from the open source community to the Qualcomm Snapdragon platform, optimized them and made them open source for all developers to use. We hope that our partner developers will be able to develop AI solutions on the Qualcomm platform that are truly applicable to a wide range of industries.

DeThings: In other words, do you or Qualcomm think that AI is a so-called "winner-takes-all" field? For example, we know that OpenAI is leading the development of a large model with trillions of parameters, so why would we need other AI in the context of this powerful AGI (Artificial General Intelligence)?

Andy Li: That's a very good question that you're asking, and I'm actually going to talk about it specifically in my talk. As you said, if AI is a winner-takes-all space, it needs to be a giant AI that can only run in the cloud, because neither the edge nor the end device can support it.

First of all, each dedicated AI model is derived from a general, large model, which we just customize and personalize based on industry needs. So we need general-purpose AI, and then we're going to break it down.

Coming back to your question, why not connect all terminal devices and industry terminal devices to the cloud? This involves the question of economic benefits. You know, in a huge data center in the cloud, the overhead is huge. Even if it's just a simple reasoning, it needs to be fully started up, consuming a huge amount of energy, both in terms of the energy itself and the energy consumption of the air conditioner in the computer room. Now, the more data centers are built in cold regions and closer to power supply areas, the better they are to reduce these expenses. That's one thing.

Second, we need to take into account that there will be more and more AI application scenarios and more and more people using it. Assuming that 1 billion people around the world use AI in the future, the energy consumed per AI launch and inference multiplied by so many application scenarios and the number of people consumes a huge amount of economic benefits. In this case, is it optimal to put all the computing in the cloud, or can we put some computing at the edge and at the end device? I think the latter is needed.

What's more, many industry end applications are more concerned about latency and latency reliability. If you go through the cloud, you tell me no more than 20 milliseconds at a time, but the latency can range from 10 milliseconds to 100 milliseconds, which is unacceptable for many industry applications. On the edge and end devices, you get deterministic and low latency. That is, the reliability of the connection.

The third point is the security of data, which is also an important reason why we promote decentralization to ensure the security and reliability of data. Of course, it's more about the security of your data. For reliability, we want data to be kept locally or within a limited range that is controllable and reachable, as a node in a larger decentralized network.

Therefore, from the perspective of the AI era, the general model may not be able to meet all specific needs in terms of economic efficiency. Moreover, in terms of reliability and security, the general large model also has some defects, and still needs industry-specific solutions. A one-size-fits-all model is not a panacea.

DeThings: We know that global mobile phone shipments are shrinking, and people have this stereotype about Qualcomm that Qualcomm's Snapdragon platform is very strong for mobile phone chips, and we want to know where Qualcomm thinks its next growth point is?

Andy Li: First of all, global mobile phone shipments, although down from before the pandemic, are slowly recovering.

Secondly, we still regard mobile phones as our main business, and we have not changed. Because many of our technologies start with mobile phones, the mobile phone market is actually the largest single-demand market ever. Although there are only a few mobile phone brands, their needs are still very similar, so the mobile phone market is a very good soil for us to cultivate new technologies in it.

In addition to this, Qualcomm is indeed pursuing business diversification. You can see we've moved into areas like IoT, XR, PCs, and automotive. In the automotive sector, for example, Qualcomm is leading the way in both advanced driver assistance systems and in-vehicle infotainment systems. We will continue to advance this diversification strategy.

The size of these emerging markets is also very impressive. In the IoT space, for example, we believe that the overall market capacity is as high as $720 billion, which is even larger than the mobile phone market. While IoT is a highly fragmented market, each industry may not have a large volume, but it has its own independent needs. But for us, we can penetrate the technical demands behind the applications in different industries, and these demands are actually the same. Our technology and patent portfolios are flexible and offer different combinations of technologies, products and solutions for different industries. I believe that this combination of technology and patents will also be very beneficial to our customers and will be able to bring them state-of-the-art technology in the most efficient way.

DeThings: In addition to the core business of mobile phones, what do you think is Qualcomm's biggest potential growth point in the future?

Andy Li: In addition to the mobile phone business, we think one of the biggest growth areas is in the Internet of Things and the automotive space.

The Internet of Things is actually one of the biggest growth points. In addition, whether it is a new energy vehicle or a fuel vehicle, we all believe that the automotive field has great potential. Even fuel vehicles have a demand for intelligence. Although new energy vehicles do not have to worry about battery problems and are easier to design as new products, fuel vehicles also need intelligent upgrades. For us, the drive form of the car is just a formality, and our goal is to make the car more intelligent, to realize the interconnection between the car and the car, the car and the transportation facilities, and more importantly, the car and the driver, which is where Qualcomm excels.

In fact, the car can be considered part of the Internet of Things because we see it as a smart connected device. Therefore, the car still belongs to Qualcomm's overall layout of the IoT ecosystem.

We believe that Qualcomm has a unique advantage in these areas because of our wide range of partners, large customer base and partnerships. Qualcomm has always been sincere in sharing our latest technology with our partners, and we are not stingy. We believe that our customers will benefit from this, and society as a whole will benefit from this.

Dethings: What do you think of the concept of Web3? Because in addition to AI, the term Web3 has been very hot in the last few years.

Andy Li: For Web3, we see it as a very new way of economic activity with potential and prospects. For Qualcomm, we play the role of an enabler in this, because Web3 is decentralized, there is no absolute big guy, everyone is an equal contributor.

As I said before, Qualcomm has been and will become more open. We're going to open up our hardware platform, and even open source our software, embrace open source software more, and we have some degree of open source code ourselves. We will empower our partners, our customers, to play a role in Web3 and possibly Web4 in the future. We provide them with weapons to wield in this vast field. Because we believe that AI is a revolution in productivity, then Web3 is about a revolution in production relations. Therefore, Qualcomm treats this emerging industry as an enabler in an open role.

Dethings: Why did you choose Aethir as a Depin platform to collaborate and experiment? It seems that very few companies like Qualcomm are involved in the Web3 space. Web3 has strong financial attributes and will involve many aspects such as token economy. Does Qualcomm have concerns in this regard?

Andy Li: Aethir is actually a very bullish partner for us. Whether it is in the field of cloud computing or edge computing, they have a long time of accumulation. If you go to Aethir's website, I think it's great to see real-time data about their resources, such as running GPUs, etc. Because they're not a conceptual thing, it's a business that really runs, and it's really put into business practice. For a pragmatic company like Qualcomm, we also want to work with such a pragmatic business partner, whether it's building a persona model or building a real production tool.

Also, we're not involved in what you're talking about in the economy, which is beyond what Qualcomm can see. We focus more on the technical aspect, we believe that technology knows no borders, and there is no good or bad technology itself. We empower technology for everyone. But personally, I don't think Web3 is a "flood beast", we need to understand and embrace all new things, and only after understanding them can we use them properly.

Qualcomm's role in Web3 is that of an enabler, and we hope that customers can use Qualcomm's technology and products to make products that are conducive to social progress and development within the framework of law and ethics.

Dethings: We've noticed that companies like Google, Microsoft, and Nvidia already have some Web3 customers in terms of infrastructure. Does Qualcomm have any further observations or attempts in the Web3 space?

Andy Li: We are open to this and welcome all customers to contact us. We believe that Web3 represents a new type of production relationship. As a technology enabler, Qualcomm is willing to share its capabilities and technology.