Linaro CEO George Grey presents HiKey 970
HiKey 970
Since the launch of Huawei's Mate 10 series and Glory V10 with Qilin 970, the AI has gradually become the standard for smart phones since 2017. The application of end-point AI has also become a hot topic in the industry. Not long ago, by CSIA-ICCAD The first international forum on artificial intelligence and semiconductor technology jointly organized by China Integrated Circuit and Moore Elite was held in Shanghai. Many experts, scholars, and industry leaders gathered at the scene to discuss the opportunities and challenges of the AI chip, focusing on the key technologies of artificial intelligence. With the future trends, AI chips and other hot topics to conduct in-depth discussion.
Huawei Zhou, director of product marketing for Huawei's wireless terminal chip, attended the conference and delivered a keynote speech. He introduced the performance of the Kirin 970 in terms of AI applications, provided platform advantages for developers, and discussed the current status and future of end-to-end AI. Here's a review of the speech below.
Those things about the application of AI chips
Application of AI chips
Currently, the industry in the AI space is broadly divided into two areas: dedicated and universal. The dedicated application scenario is very certain. The problems to be solved are also very clear. For example, city monitoring and automatic driving, AI chips in these areas are oriented toward certain scenarios. Targeted chips. The general scenario is uncertain. For mobile applications, it is not known what mobile applications will be popular next year, and the application itself is very complex. It does not solve problems through purely certain operations. Is also the difference between these two formats.
Back to the mobile phone chip, the competitiveness of the mobile phone chip is reflected in the support capabilities for unknown popular applications. The end-side AI needs to have the ability to determine the application that has not yet appeared, not to say that the current application is completed. The application is on the phone. How does the performance actually depend on the pre-judgment of the chip at design time two years ago?
Why AI is running on the end
The two core issues that endian AI solves are real-time and privacy.
For real-time demand, it is actually human instinct. Nowadays, people send out information that they hope to get real-time feedback. This demand is actually the driving force for the development of many industries, including the entire communications industry. Traced back to human beings in primitive society. Communication, At the time, people’s eyes and language transmissions were feedback in real time. As the distance grew, real-time performance began to be affected. This is also for all people, including the entire communications industry, constantly pursuing higher speeds and lower The root cause of the delay. So when it comes to the AI world, such as using a camera as a real-time application, if all the data is uploaded to the cloud and processed in the cloud and then returned, the user experience brought by this processing time is almost unacceptable. .
Another issue is privacy. Is everyone willing to send everything to the cloud? Our mobile phone is now a content production tool. It generates a lot of private information every day. It is better to put it in someone’s pocket than to put it in someone’s pocket. , This is also a very important point of the end AI.
The basic principle of AI application operation
No one application is a single load. For example, to open an application with live AI-like functionality, at least one or two threads of the CPU work at high load; just turn on the screen GPU and work on the graphics rendering class, if it is an AR application The GPU's load will be even greater; if you turn on the camera, the ISP and DSP will start to work. These components are all working. So the AI application considers not only adding a dedicated hardware unit, but how to allocate all the resources in the system. , Let the application bring a better experience for mobile users.
In the speech, Zhou Chen said that although the current AI application has not yet been widely used in terminals, there are actually many partners working together and soon there will be a large wave of AI-based applications and experiences appearing on mobile phones. Let everyone feel that the AI application can cover many areas and bring more new experiences and capabilities in the original field. There are many services underway in this area. We can all look forward to it.
Kirin 970 Raises AI Strength to Help Developers Create Innovative AI Applications
Independent AI hardware processing unit, Unicorn's 970 AI computing power significantly improved
The upgrading of mobile phone chips is a process of continuous superposition. It is best to do this with the best energy efficiency ratio without adding any area. The workload of deep learning is very high, and the performance requirements are also high. Higher. Kirin 970 chooses to join an independent dedicated hardware unit to handle AI operations. It achieves 100% AI performance by adding a smaller NPU, which is more efficient than CPU and GPU. By comparing Unicorn's 970 GPU and NPU, Obviously, independent hardware can actually achieve better results.
Zhou Chen said, 'From our point of view of making chips, improving computing power is a continuous task and it is our duty. But even more critical is that when we put our chips into place, we need developers to The AI capability becomes a user experience on the platform. So enabling developers for us is very important.
Kirin 970 provides hardware support to help developers AI application innovation
For developers, the most important are development costs, benefit-sharing and intellectual property.
The development cost is the cost of the developer's migration from one platform to another. Whether the operators are enough and whether the development tools are easy to use, and whether the documentation and support are complete, these factors determine the developer's costs to a certain extent. , Developers are concerned about the benefit sharing of applications. They may value the download volume of the promotion and application of the platform, or the reputation of the industry after the development of the APP. After that, the development model is the intellectual property of each algorithm developer. The protection of intellectual property is Very important thing. As a platform, these are issues that must be considered.
Last year, Huawei launched the HiAI mobile computing platform and opened it to developers. According to reports, the upper layer of this platform considers how 5-6 AI applications access the HiAI platform, and the lower level is the corresponding model management capability. In addition, the HiAI is offline. The mode is more suitable for the performance of the end-side AI. Load the offline model from the file. When calculating, each reasoning only needs to import and export user data from the DDR to the NPU once and it can be run. It is a very efficient mode.
It is understood that based on the HiAI V100 version launched last year, Huawei will soon introduce the V150 version, will support more frameworks and APIs, have more operators, further improve development tools, and improve documentation and support. These are all We hope to reduce the developer's cost, allow developers to debug in a familiar environment, and develop more new AI applications.
At the end of the speech, Huawei also stated, 'We hope to be able to apply the HiAI open platform to a wider range, so that different areas of different scenes can use the power of AI. In this context, we will consider launching a development board, It should be said that the industry is the first to develop AI computing capabilities as a small-end development board. It also hopes to give developers more options, and hopes to see developers gain more AI capabilities while bringing more AI applications. Landing. We are doing mobile chip platform, essentially doing well black soil, providing adequate soil, and Mobile AI should be a hundred flowers, this is what we look forward to seeing the AI industry in large areas, to everyone The real look around '.
It is reported that the HiKey 970 will be openly available for developers in mid-April to provide developers with more powerful development tools, in order to allow developers to create more new AI applications, and constantly improve the construction of AI application ecology. I believe in the near future. In the future, AI will exert greater potential on mobile phones, bringing more fresh and interesting wisdom experiences.