Shulou(Shulou.com)11/24 Report--

The 2023 World artificial Intelligence Congress (WAIC) was held in Shanghai from July 6 to 8. With the theme of "Zhaopin World generates the Future", this conference focuses on the cutting-edge technology and industrial development of AI.

The holding of WAIC 2023 this year coincides with the popularity of global generative AI large model technology, which can be seen from the theme of this conference, "generating the Future". Generative AI can be said to be the core discussion focus of this conference.

From the beginning of the popularity of chatGPT, a new wave of global artificial intelligence set off by generative AI began. According to the "Map Research report on large models of artificial Intelligence in China" released by the New Generation artificial Intelligence Development Research Center of the Ministry of Science and Technology at the end of May, 79 large models with more than 1 billion parameters have been released in China. A "hundred model war" has already begun.

The emergence of generative AI model has its inevitable reason, which is that AI, as the most important underlying technology in the digital future, will inevitably bring subversive changes to the life and production of human society. And generative AI technology seems to make people have a real experience of the concept of "AI changing the world" for the first time.

Of course, at this stage, AI still has a long way to go to really change the world, to promote the large-scale expansion and application of generative AI, but also need to be landed in the specific direction and measures.

Speaking of landing, this WAIC 2023 can be said to be very eye-catching, that is Qualcomm. This is the sixth year in a row that Qualcomm attended the conference, and their second-generation Snapdragon 8's Qualcomm AI engine won the conference's highest award, the SAIL Award (Outstanding artificial Intelligence Leader Award). Not only that, Qualcomm also brought a demonstration of its powerful terminal-side AI-enabled generative AI use case technology on the scene.

In fact, all Qualcomm's efforts have a goal, which is to depict the digital intelligence future led by hybrid AI.

Hybrid AI is the necessity of the development of AI and the key to the large-scale expansion of generative AI. There are many generative AI large model products with one billion or more parameters. These large models not only put forward high requirements for computing infrastructure, but also pose a great challenge to the existing cloud-based AI computing system.

For example, spanning AI search can provide a better user experience and search results, but each search query costs 10 times as much as traditional search methods. At present, more than 10 billion search queries are generated every day, so that it can imagine the load and cost of computing power in the cloud.

We need to find a new development model for AI that is suitable for generative AI, and hybrid AI, for now, is the best choice and can be the future of AI.

The so-called hybrid AI means that the terminal-side AI and the cloud AI work together to distribute the workload of AI computing in an appropriate scenario and time, so as to provide a better experience and utilize resources efficiently.

Specifically, in some scenarios, computing will be primarily terminal-centric, diverting tasks to the cloud if necessary. In the cloud-centric scenario, the terminal will share some AI workloads from the cloud according to its own capabilities.

Some students may think, how does the AI computing power of the terminal compare with that of the cloud? Of course, a single terminal cannot, but the number of terminals is so large that, like hundreds of millions of tributaries, it can dilute the high computing load in the cloud in many scenarios.

Just as traditional computing is evolving from mainframe and thin client to the current combination of cloud and edge terminals, AI computing is bound to develop in the direction of super-scale and complexity, so AI processing must be distributed in both the cloud and terminals in order to maximize the potential of AI.

Mixed AI can be said to be the inevitable road in the development of AI. It can greatly reduce the cost and energy consumption under the ultra-high computing demand for enterprises, not only that, but also has significant advantages in performance, privacy, security and personalization.

For example, in the hybrid AI architecture, the terminal-side AI processing has higher reliability. For example, when you use some large model products, you will often encounter slow response or even generation failure during peak hours. In the hybrid AI architecture, due to the transfer of a considerable part of the computing load to the terminal side, the cloud demand of spanning AI queries is more likely to avoid peak congestion, thus effectively reducing queuing, high delay, and even denial of service.

In terms of security and privacy, because sensitive data and information can be retained on the terminal, this is very important for both corporate and individual users.

In the hybrid AI architecture, the terminal-side AI capability is the key to enabling hybrid AI and enabling generative AI to scale globally.

In fact, before the emergence of generative AI, the processing power of AI has been transferred to the edge side terminals, mobile phones, notebooks, XR head display, cars and many other edge side terminals have also shown excellent AI processing capabilities, and there are practical applications, such as dark light shooting on the mobile phone, face unlocking and so on.

So what progress have we made in terminal-side AI?

Qualcomm, the leader of terminal-side AI, said that when it comes to the development of terminal-side AI, Qualcomm will return to Qualcomm.

As an enterprise that has been deeply cultivated in the field of AI for more than 15 years, Qualcomm has made remarkable achievements in AI-related basic research breakthroughs and large-scale expansion of cross-industry use cases.

At the same time, Qualcomm has always been a staunch supporter and practitioner of terminal-side AI.

Edge terminals such as automobiles, XR headsets and glasses, PC, and the Internet of things provide industry-leading hardware and software solutions that have a unique advantage in driving hybrid AI scale. Therefore, Qualcomm's breakthrough in terminal-side AI represents our overall progress to a large extent.

For example, we talked about the importance of generative AI, and Qualcomm's research on generative AI is actually very early, which can be traced back to spanning counternetwork (GAN) and variational self-encoder (VAE). Using VAE technology, Qualcomm created better video and voice codecs to control the model size to less than 100 million parameters.

At the software level of terminal-side AI technology, Qualcomm has been able to conduct full-stack AI research and optimization for applications, neural network models, algorithms, software and hardware, thus continuously leading the innovation of terminal-side AI experience. Specifically, in June last year, Qualcomm launched a leading software stack product specifically for edge-side AI, Qualcomm AI software stack, which can support model optimization at the software level.

For example, in the development of algorithms and models, Qualcomm has made great efforts in the development and adjustment of neural network architecture, so as to improve efficiency without sacrificing accuracy. for example, the amount of computation and delay of motion recognition can be reduced by an average of five times compared with other methods. In addition, in terms of super-resolution technology, Qualcomm's algorithm based on the Q-SRNet model, software based on INT4 quantization, and the second generation of Snapdragon 8 hardware that supports INT4 acceleration have achieved the world's first real-time super-resolution terminal-side demonstration. Compared with INT8, INT4 performance and energy efficiency are improved by 1.5 to 2 times.

In terms of quantitative research, Qualcomm has also achieved excellent results in the past few years, not only improving performance and reducing memory

It is also possible to reduce memory bandwidth consumption and save power by making the model run efficiently on Qualcomm dedicated AI hardware. At the same time, Qualcomm AI Model plug-in Kit provides quantitative tools based on Qualcomm AI research technology development, which has been incorporated into Qualcomm AI Studio.

The compiler is also a key component of Qualcomm's AI software stack, allowing the AI model to run efficiently with the highest performance and lowest power consumption. Their technical expertise in traditional compiler technology, polyhedral AI compiler and compiler combinatorial optimization AI research has achieved many advanced technical achievements.

In addition, through Qualcomm's AI engine Direct, Qualcomm is able to leverage hardware capabilities in the most efficient way, combined with the second generation of Snapdragon 8 industry-leading Hexagon processors, which will bring generative AI capabilities far ahead of this use case on the terminal.

With the enhancement of the full-stack AI optimization capability of Qualcomm terminal side, Qualcomm realized the world's first Stable Diffusion terminal side demonstration on Android mobile phones a few months ago. Stable Diffusion is an excellent generative AI model from text to image, capable of creating realistic images in tens of seconds based on any text input. It has more than 1 billion parameters and has previously been running in the cloud.

In this demonstration, Qualcomm adopts full-stack AI optimization, optimized through quantization, compilation and hardware acceleration, so that it can run on mobile phones equipped with the second generation Snapdragon 8 mobile platform, and can perform 20-step reasoning in 15 seconds to generate a 512x512 pixel image. This is the fastest reasoning speed on a smartphone, comparable to cloud latency, and user text input is completely unrestricted.

Not only that, Qualcomm also completed the ControlNet terminal side demonstration on the world's fastest mobile phone. The ControlNet image generation image model is a language-visual model (LVM) with 1.5 billion parameters, which can be more accurately controlled by adjusting the input image and input text description.

In this demonstration, ControlNet can interoperate efficiently on the terminal side, and through a set of full-stack AI optimization across model architecture, AI software and neural network hardware accelerators, 16-step reasoning can be completed in 12 seconds to generate AI images, providing an efficient, interesting, reliable and private interactive user experience without visiting any cloud.

‎ at the WAIC scene this year, Qualcomm brought back the technical presentation of these two generative AI use cases, which has become one of the most eye-catching scenery in the exhibition.

Next, Qualcomm is planning to support models with tens of billions of parameters on the terminal side in the future, which will become a major differentiation advantage of products based on Qualcomm technology.

Hardware to ecology, hybrid AI in front of accelerating landing, we talked about Qualcomm's full stack optimization capability at the terminal side AI software level. Its advantage is that once the model is developed, it can be used in different places, and then combined with hybrid AI deployment to form a killer combination, which will help generative AI scale expansion on different terminals and achieve the popularity of generative AI.

With a strong software foundation, Qualcomm will be able to promote the popularity of hybrid AI architecture at the terminal level through actual products.

Specifically, Qualcomm's hardware provides industry-leading energy efficiency, nearly twice as much as mobile competitors. Among them, Qualcomm AI engine is composed of several software and hardware components, which can realize terminal-side AI acceleration on Snapdragon and Qualcomm platforms.

At this year's WAIC conference, the Qualcomm AI engine of the second generation Snapdragon 8 mobile platform won the top award of the 2023 World artificial Intelligence Conference: the SAIL Award (Outstanding artificial Intelligence Leader Award).

The AI engine of the second-generation Snapdragon 8 platform brings a 4.35x improvement in AI performance and a 60 per cent improvement in energy efficiency compared to the previous generation, providing a strong performance foundation for more and more innovative AI use cases and AI-enhanced user experiences.

Qualcomm AI engine uses heterogeneous computing architecture, including Hexagon processor, Qualcomm Adreno GPU and Qualcomm Kryo CPU, all designed to run AI applications quickly and efficiently on the terminal side. Through heterogeneous computing, developers and OEM vendors can optimize the AI user experience on smartphones and other edge-side terminals.

Among them, Hexagon processor is the most important part of Qualcomm AI engine. In the second generation Snapdragon 8 mobile platform, the latest Hexagon processor has a series of innovations, including providing dedicated power supply system, supporting microslice reasoning, INT4 precision, Transformer network acceleration, etc., which can be combined with Qualcomm AI software stack and AI Studio to provide full-stack AI capabilities and optimization means.

In short, Qualcomm AI engine is the core of Qualcomm's terminal-side AI advantage, it plays an important role in Snapdragon platform and many other products, it is the crystallization of Qualcomm's full-stack AI optimization for many years, and can provide industry-leading terminal-side AI performance with very low power consumption, which I believe everyone is familiar with.

In addition to the Qualcomm AI engine, another thing that can not be ignored is that the scale of the edge-side terminals deployed by Qualcomm is very large, the number of listed user terminals equipped with Snapdragon and Qualcomm platforms has reached billions, and hundreds of millions of new terminals are still entering the market every year. These terminals cover a wide range of products, including mobile phones, cars, XR, PC and the Internet of things and so on.

Behind the huge edge-side terminal network, Qualcomm has the advantage of ecological scale expansion of hybrid AI industry around the world.

In terms of the most basic phones, Snapdragon Mobile platform is the leading mobile platform to enhance the top Android AI experience, including more than 2 billion AI-capable processors that have been shipped. Snapdragon is also a leader in the mobile platform AI benchmark, such as the top 20 in the industry's well-known AI Benchmark.

In the automotive sector, for example, Qualcomm is also a leader in cockpit and in-vehicle infotainment solutions. at present, all major automakers around the world have adopted Snapdragon cockpit platforms to enable their digital cockpit systems. Many automakers, including Honda, Mercedes, Renault, Volvo, BMW, General Motors / Cadillac, Great Wall Motor, Toyota, Xiaopeng Motor, GAC GROUP and others, have launched mass production projects or are designing platforms with Qualcomm solutions to enable the industry-leading in-car AI user experience to land quickly in a safer, more comfortable and reliable way.

In the field of PC and tablets, Snapdragon's computing platform integrates Qualcomm's AI engine to support powerful terminal-side acceleration, bringing better quality, performance and efficiency to the latest applications. And in the field of XR, more than 65 XR terminals using Snapdragon platform have been released so far, including many popular terminals from brands such as Meta, PICO and Lenovo, which integrate Qualcomm terminal-side AI and Snapdragon Spaces technology to provide a more immersive experience and better adapt to the world around them.

In the broader field of the Internet of things, Qualcomm has deep cooperation with more than 16000 customers across different vertical areas. Massive end products with AI processing power embedded in high-access Internet of things chipsets and platforms support terminal-side data analysis (such as video) in an efficient and feasible way, driving innovation and transformation across multiple segments, including robots, smart cameras, retail and urban infrastructure.

It can be seen that Qualcomm is not only leading the basic research at the forefront of AI technology, but also actively cooperating with Qianhang Baiye through full-stack software optimization and mature hardware products to carry out cross-terminal industry expansion and ecological empowerment, so that the edge of intelligent network connection driven by hybrid AI can land at a faster speed.

Conclusion maybe we can imagine the future of large-scale expansion of generative AI, using mobile phones will become more and more simple and intelligent, and it will become a real personal assistant.

At the same time, PC notebooks will become a more powerful productivity tool, and with the powerful generation ability of AI, our daily workflow will be greatly accelerated, and tasks that used to take hours or days can be completed in just a few minutes.

In addition, generative AI will also make the car a highly personalized experience space, which will help you plan your travel route, provide customized experiences and content such as music and podcasts, or list today's work items on your way to work.

These future sights need hybrid AI architecture to become the underlying driving force, and this trend is unstoppable. Qualcomm is making hybrid AI a reality by virtue of technological innovation, global scale and ecosystem empowerment.

I believe that by then, we will be able to experience firsthand how hybrid AI subverts the way we live, work and play.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.