With the development of technology, artificial intelligence gradually penetrates into our life and completely changes our world, and the smartphone circle has set off a wave of artificial intelligence processors, which can be said to be a major new round of smart phone revolution.
For now, Huawei and Apple have taken the first step in the field of smart phone AI chip.China's latest Kirin 970 processor has a dedicated NPU, and used in the just released Huawei Mate 10 series and glory V10 , Making it a 'real artificial intelligence smartphone.' Apple named its newest processor A11 Bionic, mainly because of its built-in artificial intelligence 'neural engine.' The two companies walked in the smartphone AI chip The forefront of the application.What is smart phone AI chip 'Ssangyong' strategy and layout, this article to be a careful sort.
Unicorn 970 Vs. Apple A11 Bionic who better?
Huawei Unicorn 970 processor first released in the A11 Bionic Apple, becoming the world's first with a separate NPU dedicated hardware processing unit of the mobile phone chip, and the first use of the chip Huawei's mobile phone Mate 10 was also released, indicating that China's artificial intelligence chip applications In a leading position.Then these two have the birth of the AI phone chips What are the similarities and differences?
First of all, from the AI module part, Huawei and Apple have chosen the ASIC way to achieve.In order to meet the needs of different AI scenarios, the current AI chip is divided into several schools: parallel computing power and high power consumption GPU, an FPGA that redefines the internal circuitry, and a deep custom ASIC. Each of these technology genres has its own advantages: in contrast, ASIC accelerators tend to outgrow their counterparts with the same process FPGAs Accelerators are 5 to 10 times faster and far superior in power consumption to GPUs, and once mass-produced, the cost of ASICs will be much lower than that of FPGAs. Both Huawei and Apple agree at this point to choose between The ASIC way, we can see both planning, vision are the same, but also earlier than peer force AI.
Second, in terms of performance, Kirin 970 HiAI architecture for the first time integrated NPU dedicated hardware processing unit, bringing advanced AI computing capabilities, so that smart phones really enter the AI era.Apple also uses a dedicated neural network engine (Neural Engine) , But not a dedicated NPU. In contrast, for the first time, Huawei integrated the hardware computational acceleration capability of the neural network model into the chip at the mobile device level, making it the first time in the industry that a mobile-phone chip specifically designed for artificial intelligence Processing unit. The HiAI architecture maximizes CPU / GPU / ISP / DSP / NPU performance with more efficient and flexible heterogeneous computing with significantly faster performance and energy efficiency than CPUs and GPUs. For example, Module, unicorn 970 can identify 2005 pictures per minute, and if the CPU only 95 per minute, the former than the latter 20 times.
The same is the artificial intelligence chip, Kirin 970 and Apple A11 performance? Still in the picture recognition test as an example, Kirin 970 can identify 2005 per minute, while the Apple iPhone8plus 899, the unicorn 970 performance is Apple A11 two Times the number of this one may be a bit single, we recently released from China Mobile's "China Mobile 2017 Terminal Quality Report" can see a more comprehensive evaluation.In the report, Huawei mobile phone performance in the machine, chip, camera Such as multi-dimensional performance are excellent, won two in one fell swoop.Among them, the flagship business of Huawei Mate 10 Pro occupy the "cell phone comprehensive evaluation list" more than 3,000 yuan price segment ranked first position.Kylin 970 is based on three full marks Won the chip communication performance evaluation, AR performance evaluation, AI performance evaluation of the three first, beyond the peer Apple A11, Xiaolong 835, MediaTek X30, Spreadtrum SC9853 and other platforms, called the evaluation of the 'strongest chip' As you can see from the report, Kirin 970 outperformed Apple A11 in both performance and energy efficiency aspects of AI.
Third, from the point of view of chip integration, Kirin 970 uses the latest TSMC 10nm process, integrating 5.5 billion transistors.Apple A11Bionic also uses TSMC 10nm FinFET process, integrates 4.3 billion transistors.But it is only AP, if Together with the roughly 1 billion base-band transistors (based on the specification-leading Qualcomm baseband X16), there are 5.3 billion transistors, which are not much different from the number of transistors, but the Kirin 970 has more features. For example, in terms of communications, this is Huawei's housekeeping skill. The Kirin 970 implements the Cat18 / Cat13, the world's most advanced standard, and uses the most advanced technologies such as 4 * 4 MIMO, 5CC CA and 256QAM to maximize fragmented spectrum aggregation Bandwidth and Aggregate Peak Capability up to 1.2Gbps download rate, once again refreshed the fastest speed of mobile Internet connection; and iPhone8 and iPhone X communication specifications are unknown, some say Cat9, the peak download speed is 450Mbps. Even if Qualcomm is currently the most advanced commercially available baseband X16, Cat16 only, the peak download speed is 1.0Gbps. Meanwhile, the unicorn 970 commercial dual card dual 4G dual VoLTE, the successful realization of a On board two SIM cards support 4G VoLTE call, provide the main and sub card consistent high-definition voice, video call experience. This aspect should be better than Apple's Kirin 970 A11.
Therefore, in comparison, both an area of considerable, the same number of transistors, Kirin 970 better performance, better experience.
Closed or open, which is more suitable for AI applications landing?
In terms of application and ecology, Kirin 970 and Apple A11 Bionic, which belong to the Android camp and the iOS camp, also have their own characteristics.
Applications, Apple's AI technology is more cool, mainly used for Face ID, scene recognition when taking pictures, and Animoji and other applications, the application of a narrow surface, and application development environment is relatively closed.Wuhan is more emphasis on pragmatic, application scenarios But also relatively more rich.After joining the HiAI mobile computing framework, so that the traditional photography into the smart photography stage, the more you use the more clever.For example, AI Hui eyecare, with the powerful computing power provided by the AI chip, the AI smart camera function can automatically Food, plants, stage, pets, portraits and other 13 kinds of scenes to provide exclusive optimization of the program.In voice, the use of artificial intelligence hardware and software combined to achieve real-time text-to-interactive translation and knowledge map translation, but also easily Face to face speech translation, and acceleration through the AI chip has the ability to learn, the more you use the more clever ... Equipped with the unicorn 970 glory V10, the use of powerful AI computing power to achieve a face unlock, and richer AI applications.
Speaking of openness, we can not but talk about the ecosystem, and in that regard Apple introduced Core ML, a machine learning framework that lets developers use AI models in a variety of formats in the simplest possible way, but Core ML Not open-source, and the open ecology is very important, so that more developers involved in it faster to make the entire mobile AI ecology continues to evolve and improve.Apple ecology + end way may only be suitable for Apple itself.
Restricted by the Android operating system, Huawei is actively trying to make up for the lag of iteration of the Android version, although it is not eco-friendly, including launching an open AI environment that supports both Tensorflow Lite and Caffe More mainstream neural network programming framework to facilitate developers to better call NPU in the application.
In addition, Kirin 970 chip release, it has announced that the unicorn 970 chip will be based on artificial intelligence computing platform open to all partners, so that every partner can enjoy the convenience of the end-side artificial intelligence to develop A better application of Huawei after the establishment of an open AI ecosystem, more outstanding developers, the third-party industry will use the platform to create more new applications, users will be brought more valuable AI Application, constantly rewrite the Android mobile phone in the eyes of the user's image.
In the meantime, Google will strengthen Android's support for AR functions next year, and the new version of Android framework will also update the AI part, so Android camp will not lag behind too much.Huawei is backing Android and a more open ecosystem will have more outstanding developers And third-party applications to promote AI applications landing.
AI mobile phone chips which way to go?
Although Huawei Kirin 970 and Apple A11 Bionic have played a role in the application of AI chips, the innovation in this field involves artificial intelligence algorithms, programming languages, computer architectures, integrated circuits, and semiconductor technologies. It is by no means accomplished in a single action Things.
Followed by Huawei and Apple, more handset chip makers have also been involved in the development of AI. Qualcomm neural network processing engine (NPE) based on distributed architecture, you can run in the snapdragon heterogeneous platform CPU, GPU, DSP and other units, AI is expected to be included in the high-end SoC products in the coming months.Compared with Huawei and Apple's ASIC, Qualcomm believes that flexible AI hardware is the better choice at present, The combination of DSPs and GPUs is even better.Another major handset chip supplier, Samsung, is also working on a dedicated artificial intelligence chip, the biggest difference between these chips and current AI solutions is their off-line data processing capabilities, namely device data The device itself can handle these issues without having to fully import it into Samsung for processing and return. Samsung's view is that using current CPUs in smartphones to handle deep learning algorithms is inefficient and adding NPUs to be the future In addition, it is said that MediaTek has also completed the design of the built-in AI computing unit of the mobile phone chip. Several major chip giants work together to force the AI phone chip, AI chip into the phone's popularity. Counterpoint market research firm predicts that by 2020, equipped with independent AI chip smart phones will exceed 1/3 of the number of over 500 million units.
With the development of diversified AI chip market, but also a variety of bottlenecks to be solved.Current technical challenges lie in increased power, algorithm optimization, lower power consumption and better integration, the future will encounter security Problem An important measure of the computing performance of an AI chip is called computational effort, which is commonly referred to as FLOPS, as a measure of the number of floating-point operations performed per second (also known as peak speed per second). The already cost-effective SoC in cell phones, adding NPU or other AI modules, poses a greater challenge to the complexity of the chip architecture design, the cost and the software architecture brought by system integration.
Security and privacy are also inevitable problems.At present, many terminal data are sent to the cloud for analysis and treatment, the recent edge computing began to rise, this part of the task on the side to complete, can greatly reduce the risk of data leakage and hacking.
The biggest challenge is also how the AI chip is really utilized. Now the phone has an AI chip, but want to really rely on the developer to use. Gartner's technical analyst Anthony Mullen said: 'These technologies are not currently Suitable for small businesses because manufacturers tend to work with large, third-party companies until such complex hardware is developed. "For some small and medium-sized developers, is it affordable for development? Time and cost are still not clear.
However, it is delightful that Unicom 970 is a brave attempt by Huawei in the field of artificial intelligence, and in the future Huawei AI series chips will continue to improve performance in terms of manufacturing process, chip architecture, resource allocation and algorithm optimization. Step, will undoubtedly completely change the AI industry and the mobile phone chip industry play.A brief few years, Huawei baseband baseband chipset quickly surpass Qualcomm, and in the application of artificial intelligence leading apple, the future in the smart phone technology innovation, Who can dominate the market, but also wait and see.