Live: Huawei unveils AI chip and computing framework 华为人工智能（AI）芯片与全新开源计算框架发布会

Ladies and gentlemen, friends from the press, analysts, and friends, good morning, good afternoon, because we have more than 100 friends from the press joining us online. They come from Europe, so good morning to you. Many of you may have attended the Huawei Connect conference last year in Shanghai. At that event, I officially announced Huawei's AI strategy as well as our full-stack, all-scenario AI portfolio. I also outlined major changes that we need to make to help make AI more pervasive and accessible. We hope all industry players will work together to drive changes and close the gap between AI-related reality and AI-related potential. And over the fall, we have been working hard on several of those areas. So today, the launch event is about an update of our latest progress. First of all, please allow me to brief you on our AI strategy, which has five pillars that I introduced last year. First, we invest in AI research. We set up a Noah's Ark Lab. We develop fundamental ML capabilities in computer vision, natural language processing, and decision inference, focusing on data and power efficiency, namely using less data, computing, and energy. Security and trustworthiness is another area of focus. Automation and autonomy is the third one. So that's the responsibility of our Noah's Ark Laboratory. The second part of our strategy is to build a full-stack AI portfolio adaptive to all scenarios, including both standalone and cooperative scenarios between cloud, edge, and the device. We also aspire to provide abundant and affordable computing power and an efficient, easier-to-use AI platform with all four pipeline services. The third pillar is to cultivate a challenging and open ecosystem, working extensively with global academia, industries, and partners. Fourthly, we will strengthen the existing portfolio by bringing AI into all of Huawei's products and solutions to create greater value and make them more competitive. The fifth part of that is to drive operational efficiency by using AI to automate high-volume, repetitive tasks for better efficiency and quality.

Last year, I also launched our full-stack, all-scenario AI portfolio. By taking today's opportunity, I'd like to reiterate and take you through this full-stack, all-scenario portfolio. The portfolio covers all development scenarios, including public cloud, private cloud, edge computing, IoT, industry devices, and consumer devices. That's what all-scenario is about. And the full-stack is from a technical or functionality perspective, that includes chips, chip enablement, training and inference frameworks, as well as application enablement. There are a couple of areas for full-stack. One is the Ascend IP and the chip series, which is based on a unified, scalable architecture. In this series, we have Ascend 910, Max, Mini, Lite, Tiny, and Nano. The other part is the CANN chip operators library and highly automated operators developing toolkit. The third part is MindSpore, a unified training and inference framework for device, edge, and cloud. The other part is application enablement, which is ModelArts, including full pipeline services, hierarchical APIs, and integrated solutions. So last year, we launched our AI strategy as well as the full-stack, all-scenario AI portfolio. Also last year, we launched our very first AI processor, the Ascend 310, primarily used for edge computing and inference. Today, Ascend 310 has been widely adopted in a range of edge scenarios. There are three primary series based on Ascend 310: MDC, Mobile DC; the other is Atlas, including as well as computing AI accelerated modules and servers; the other is cloud services. Today, these solutions are widely used in our partnership with carmakers to address intelligence and automation needs. The Atlas cards and servers are widely used by our partners, and they use them for smart transportation, smart power. It's also used widely for heating and wastewater processing that is closely linked with people's livelihood. Cloud services based on Ascend 310 are for video analysis. There are more than 50 APIs related to this. Based on Ascend 310, providing cloud services, the daily API calls has exceeded 100 million, and this figure is expected to hit 300 million by the end of this year. At the same time, we also launched ModelArts, a full-pipeline model production service. It provides model developing services spanning the full pipeline from data collection, model development, to model training and deployment. Today, there are more than 4,000 training tasks per day for a total of 32,000 training hours. 85% are related to visual processing, 10% audio processing, and 5% are related to machine learning. And today, there are more than 30,000 developers using our ModelArts. Today, what I'm going to show you is Ascend 910, the industry's most powerful AI processor.

In October last year, we actually disclosed the tech specs of Ascend 910. Today, I will show you more about how it actually performs in tests. Test results show that the Ascend 910 processor delivers on its performance goals. For high-precision floating-point operators, Ascend 910 delivers 256 teraflops, and for integer precision calculations, it delivers 512 teraops. More importantly, its max power consumption is only 310 watts, much lower than specified. Under the same conditions, its computing power is twice that of the mainstream benchmark. Ascend 910 performs much better than we expected and better than our design specs. It is now already used for full AI model training. In a typical training session based on ResNet-50, the combination of Ascend 910 and MindSpore trains models about two times faster than other mainstream training cards. Using TensorFlow, Ascend 910 can train 1,803 images per second, while the existing benchmark is 965 images per second. As for the key technologies, we have a video clip. This is a 7nm Da Vinci architecture-based AI core for the compute engine. In addition to scalar and vector units, the AI core integrates a 3D cube computing engine. It completes 4,096 MAC ops per cycle, more than two orders of magnitude larger than what CPUs and GPUs can deliver. Ascend 910 has 32 cubes inside, providing 256 teraflops. Not just a powerful AI coprocessor, Ascend 910 is a highly integrated SoC comprised of CPUs, DVPP, and task scheduler. As well, Ascend 910 has self-managing capability that can minimize data interchange with the host. By eliminating this overhead, the unprecedented computing power can be fully utilized. Highly efficient communication mechanism is another key for a training system. Ascend 910's self-developed HCCS can deliver 240 gigabits per second for each port. Latest PCIe guarantees twice the throughput compared with previous generations. Lastly, on-chip 100G RoCE enables direct data exchange between multiple nodes, which improves training cluster performance and flexibility. High computing power, high level of integration, and ultra-fast interconnection together build Ascend 910, the world's most powerful AI processor.

Ascend 910 is only the start. Moving forward, we continue investing in AI processors to meet different needs of a broad range of scenarios. There are a couple of Ascend series for edge computing based on Ascend 310. Our plan is to launch Ascend 310 in 2021. The existing MDC is based on Ascend 310, currently used for autonomous driving. In the future, for large-scale commercial use, we'll be launching Ascend 1616 and Ascend 620. For AI training, today we officially launched Ascend 910. In the future, we are going to launch Ascend 920, so that we can come up with a whole series of AI processors for different scenarios, so as to provide powerful and affordable computing resources for research and industrial adoption. Today, I would also like to announce the release of MindSpore, our all-scenario AI computing framework. We know AI computing frameworks are critical to making AI application development easier, making AI applications more pervasive and accessible, and ensuring privacy protection. That's where an AI application development framework comes in. For that purpose, at Huawei Connect last year, we announced three development goals for our AI framework: easy deployment, which substantially reduces training time and costs; efficient execution, which uses the least amount of resources; and most importantly, it needs to be adaptable to all scenarios, including device, edge, and clouds. MindSpore marks significant progress towards all three dimensions. It is adaptable to all scenarios across all devices, edge, and clouds, and provides on-demand cooperation between them. Its 'AI algorithm as code' design concept allows developers to develop advanced AI applications. That means people working on algorithms do not necessarily have to have strong coding capabilities, so that would bring much ease of use and train models much more quickly. Through framework innovation as well as co-optimization of MindSpore and Ascend processors, our solution can ensure stronger performance and more efficient execution. MindSpore does not only support Ascend 910, it also supports other GPUs and CPUs or other processors available in the industry. Many people have asked me the question: now that we have TensorFlow or PyTorch, what's the point of MindSpore that Huawei works on? I've been telling these people that right now, none of the existing frameworks can support all scenarios. Well, you can look at our portfolio, we cover devices, edge, and clouds. Furthermore, data privacy protection has become a more important topic than ever, and the support for all scenarios is essential for enabling secure, pervasive AI. This is also a key component in our MindSpore framework. Resource budgeting environments can be big or small as needed. MindSpore also helps ensure user privacy because it only deals with the gradient and model information that has already been processed. It doesn't process the data itself, so private user data can be effectively protected even in cross-scenario environments. In addition, MindSpore has built-in model protection technology to ensure the models are secure and trustworthy. In addition, MindSpore is built on a concept called 'AI algorithm as code.' This design concept allows developers to develop advanced AI applications with ease and train their models more quickly. In a typical neural network for natural language processing, MindSpore has 20% fewer lines of code than existing frameworks on the market, and it helps developers raise their efficiency by at least 50%. Through framework innovation as well as co-optimization of MindSpore and Ascend processors, our solution can help developers more effectively address complex AI computing challenges and the need for a diverse range of computing power. This results in stronger performance and more efficient execution. In addition to Ascend processors, MindSpore also supports other processors as well. And we also make MindSpore developer-friendly. Adopting an 'AI algorithm as code' concept, MindSpore enables a host of new technologies, including source-to-source automatic differentiation, which vastly outperforms graph and operator overload. MindSpore can achieve differential expression and compiler optimization for any operator and automatic generation of inverse operators. This makes model development as easy as ABC. As data sets and model scales grow ever larger, parallel processing is inevitable. Regardless of how time-consuming and difficult it is to implement, MindSpore is here to help. By simply defining a standalone model, MindSpore can automatically implement multi-machine hybrid parallel operations. MindSpore also supports both static and dynamic graphs, allowing for seamless switching with only one statement, making model debugging easy and efficient. MindSpore provides Ascend-native runtime efficient technology to maximize the computing power of Ascend chips. In the host-device model, interaction between the CPU and GPU creates large memory and data overhead. MindSpore controls and executes the entire neural network model training on the chip, reducing interaction with the host CPU and improving training speeds. Existing distributed training models introduce central controls to define gradient synchronization points. MindSpore instead uses point-to-point distributed gradient aggregation and completely eliminates control overhead. Software and hardware are optimized to map different types of operators to their optimal computing unit and data layout, achieving exceptional performance. With MindSpore, mathematicians and researchers are being given an invaluable tool that makes theoretical exploration and innovation easier and more efficient. To drive AI application, in order to encourage the entire industry to build MindSpore into a real all-scenario computing framework, MindSpore will go open-source in the first quarter of next year, so as to help every developer and encourage every developer to participate. With this launch of Ascend 910 and MindSpore, Huawei has unveiled all the key components of our full-stack, all-scenario AI portfolio. As you might be aware, Kirin Lite is already used in Kirin 980, used in Honor and Nova phones. And Ascend Tiny, we're going to have Kirin 990, which Richard Yu will be launching at IFA this year. Kirin 990 is based on a neural processing unit, Ascend Lite. So last year, we promised a full-stack, all-scenario AI portfolio, and today we delivered. This is a new milestone for Huawei, it's also a new start. We look forward to working together more deeply with partners around the world to build pervasive AI to benefit every individual, every home, every organization. We also hope all partners can work together with Huawei to promote the development of AI so that AI can benefit society. Thank you.

I would like to address your question with a kind of commercial or advertisement. At Huawei Connect this year, scheduled on September 18th in Shanghai, we'll be launching more new groundbreaking AI products. Ken, who will be serving with you, more detailed information by then. As for the next steps for our plan for the AI industry, we'll be promoting and working on our already established AI strategy and our full-stack, all-scenario AI portfolio step by step. You know, all the key components have already been launched. We'll continue to work on the constant iteration and new versions. We also hope our AI-related products and solutions can gain more adoption and recognition in the market from customers and partners, so that AI can benefit more people in China and around the world.

Our AI strategy is not necessarily tied to 5G. 5G is 5G, and AI is AI. As for our AI strategy in India, it's no different from our overall strategy. We are going to develop Atlas cards, servers based on Ascend processors as well. We're also going to work on MDC products based on the Ascend processors as well. Those products can be provided to universities, partners, and enterprises in India as they develop industry AI applications to address industry-specific issues and challenges and to create value. Similarly, we have training and inference cloud services based on Ascend processors as well that can also be made available in the Indian market. Universities, research organizations, and enterprises can use those AI cloud services for their research and application development needs, so as to address the requirements and challenges of different industry verticals. In short, we are going to provide products and services, and we are going to work together with partners to address the needs of different industry verticals, serve them, help them be more efficient, so that we together can contribute to the development of society.

I think the incident on May 16th has had no impact on the execution of Huawei's AI strategy and also the R&D and commercialization of AI-related products. Even though there have been many changes in the external environment, the execution of our AI strategy, the R&D and commercialization of Huawei's AI products have been going steadily and orderly. Also, we are living up to our original plan and expectations. As for autonomous driving, since we launched our joint research results with Audi, we have stepped up our investment in that area. We also work together with all carmakers on MDC products. We support and help them to come up with autonomous driving solutions. Of course, there is still some way to go before we reach fully autonomous driving, but I think it is possible that we can get there step by step. We also develop autonomous driving solutions ourselves, which we take and work together with our car OEMs. So for those co-designed cars, we can demonstrate the autonomous driving capabilities and experience. When we look at the future of autonomous electric cars, we position ourselves as a provider of incremental components. For that purpose, we have set up a new business unit called Intelligent Automotive Solutions, so as to drive the progress of this business. Our vision is to leverage our AI capabilities in our full-stack, all-scenario AI portfolio, working together with the carmakers to really deliver on autonomous driving.

I would say we are satisfied, because if I say we're not satisfied, many people would have a different opinion on that.

We all know the development of an ecosystem is very important. That also points to the biggest challenge for MindSpore: whether we can set up as compelling an ecosystem as the other mainstream frameworks currently available in the market. But there are advantages for Huawei when it comes to ecosystem development, because we have smart devices in our portfolio. We have a HiAI engine in the smart devices, where the developers on mobile applications can access Huawei's AI computing power to develop different types of applications and services leveraging our capabilities. In addition, we also have computing power. The combination of our computing power, or the processor capability, with the MindSpore computing framework can help Huawei do something where the other framework players cannot. As I mentioned just now, through architectural innovation of MindSpore and through co-optimization of MindSpore and Ascend processors, it can substantially increase the efficiency of AI. Certainly, we have had a mission to ourselves: we want to make AI development as simple as possible. If you look at the reality today, AI is still a game for experts. As I mentioned last year, we have ModelArts, our full-pipeline model production service. The idea is to enable ordinary engineers to be able to develop AI applications. If ModelArts can truly deliver on that mission, then we would contribute hugely for the development of the ecosystem. As we said last year, whether AI can truly take off, there are a couple of important factors. One of which is whether the enterprise market would take off, or in other words, whether there would be more enterprises and industries to adopt AI.

Huawei has quite comprehensive and a strong enterprise business that's also another advantage we have as we built out our AI related ecosystem.

We are going to rely a lot of our advantages of a comprehensive portfolio spanning device, edge, and cloud, and also our full-stack, all-scenario AI portfolio to address that challenge head-on. And to work on ecosystem, and only when the ecosystem truly takes off, the all-scenario full-stack portfolio can deliver value.

Just as long as you in court even see cigar synth enjoy a lean quiz i invoices are kuma report which with a heart attack is muscle teacher watch romantic she wanted in potential now is general Hammond took a sample first wanna wear one the Air Jordan wouldn't take TV and enjoy identical atras take a Bangkok Hotel Fuji TV the sand enjoyed in the UFO I bound to the in Connecticut Austria Quantico Jason a NGO her in Haifa so treaty of ghent I hope we can have a Ascend 910 based cloud services launched in the UK and I have not thought about a specific timeline for that but I do believe Ascend 910 based Atlas series for both accelerator cards and the service as well as Ascend based cloud services can truly deliver value to help universities and research organizations to do AI related research or work on AI related applications.

Kansai Eric $80.00 essentially Ohio Latika tiny now coz any question here was IP who eats a young a champion tired Lucy now the are gotta hot sentence orange mr. Horwitz I realized some fateful the girl nano level when necessary to anticipate the waxing that you never to guide the Sun King Hamad it's right honey free tire woman she didn't took a Joe Jordan thumb because in July the lettuce well I imagine she didn't Joe Jordan couldn't Siddhant and I'm currently our plan is to use Ascend Nano for smart wearables. Right now there is no product launched that is based on the Ascend Nano IP and for Ascend Tiny it will be used in Kirin 990 and accordingly in a whole range of smart devices powered by Kirin 990.

We're going to have cloud services for both inference and training based on Ascend 910. Currently we already have two international nodes, one is Singapore, the other is Hong Kong that can cover Vietnam. The other words we can have cloud services based on Ascend 910 already covering Vietnam market.

Woman cancel by the integration via the complete agree that the recipient the girl me santa casa da SE dabangg cappucci ho young for the function nominations system our strategy our strategy is and will be the same. Ascend series of processors will not be provided to the market standalone rather they are going to be embedded into our acceleration cards or service or clouded services.

In addition we also are open to discuss partnerships with the AI chipset developing companies so that there are chipsets of various kinds can be possibly used for different types of computing scenarios. This is so bad yeah the team can do you mean just the top person pushing do that therefore it's not going to happen where AI chipset will be positioned as a standalone business and is not a new stage from that point of view.

Like a language in yesterday's ago phones Ian renovations and Joe cheetah so yo you make order tickets to teen wow it's a new sajida take a shell show surely our champion cannot wait Shah Baba the musician not always in sahaja yoga kshema senior have symbiote has normal white it seems is not as pessimistic as a mr. when I said when Dunston you say Ben this is big a big wonder you take English Indian gotta be danced Ayutthaya hot at all maybe that was a forecast and your hospital worse the scenario worst-case scenario now the situation is much better than what was forecasted the yes I'll go buy a toy images your dad but it could happen that the business could be more than 10 billion US dollars less she'd unexpected engines and sometime and that's only for smart devices.

The price will not be very high but is not decided. Let me clarify the price point for Ascend 910 will not be higher than those offerings but is not finalized yet.

Alive solutions and beliefs Italy 20 since I aha Otto oh wait Tom Jaffa hon chan hak-kan AI Shangguan de take a total middle 20 Tanzanian soo - aha yes yo a either take a toll on me T'Challa you see the dirty dog answer change nah way it's such a famished the makanda now how's that work area mean hi AHA is using the AIT Sri aha it's a makanda trigger area that she should read the wiki so I said take positions a idle only when he announces or cigar CCTV falender Cynthia photos Alyssa with the chief of our adult Indian can get easily the third event he so you could have a tiny woman 3 thou omen as adult I take a family and say she de Noli woman fit and talented on lowly on search engines also visited commercials for the TGA CEO Joe I think the discussion around ethics does not only apply to AI as society proceeds especially on the technical ground as we are embracing information society technical ethics is a topic of huge controversy. We also have noticed the cost in the efforts of Europe in this area and we highly appreciate those efforts and Huawei would follow and support related to the guidelines and recommendations.

And we also know that ethics related to technology or AI would require more discussions and more enhancements. Lou Gorman TN and Elijah Cohen chiseler need a venti Pakistan guten on the lulav NT in decision a transgenic when she when she simply put the technical or AI related their ethics are voiced down are to you the relationship between human and in nature today.

In ancient times when humans our strengths were much fewer than that of nature and then we felt when it comes to technical ethics is about gaining strength from the nature and the respecting the nature. Co a gumiho Linda Lillian toggle of the land leader after the Industrial Revolution human power exceeded the power of nature. Omaha jisu the religious to the daughters and the other Tamiya Univision and then the emphasis for moral ethic surrounded technology was around how to make it more favorable for human society through Nevada McConnell attend or take a dollar per head of an invention take taeho bein an event he managing poha when he eats chicken show the chisel Alicia without even this is Walker search engine is the gender ratio confident.

Then she put on the gaudiya Fadden tion patron farinelli's patron van ho ho the derry-o recession su chio put on the one day the material only overlay in dungeons and adult high is indeed in ship syndrome Concordia put on the shelter but different countries have different stages when it comes to their economic developments their history and their culture therefore they may come up with a differently understanding around what is harmonious coexistence that is a situation and a reality that people have to very mind any respects when we have a technical ethics our discussions bearing in mind the differences of a timing of a geographies and of a specific specific situations Peru languages and legal treatise in Union cities and in Syria Caillou is sitting the Genco nabatea a girl sitting the universe acadia soda chin so young boy was Doron DiSanto pesticide prokovich adequate as an amputee so that tech boom gua is a crazy soda so he chooses high they don't hire a book to give you one example a very typical AI application is our computer vision and in a video surveillance surveillance this is something that is accepted in the UK it is a set there are more than 500,000 cameras on the street in the UK is also accepted here in China but it's definitely not something that people accept in Germany or other European countries that's a very good example in terms of the differences that people should bear in mind when we talk about technology ethics.

Well mr. Dooley the Gherkin number and Tiernan's happened that doesn't get ya yeah I shall go I would say we're being working hard to see whether we can develop or light up this a beacon of talents but we need to take away Tennessee approach to see whether we can really deliver the data.

We have done one thing recently that some people in the industry may not necessarily understand. We made the compensation on members of the top of PhD students widely available just a woman today take sound gential gafsa woman thundermans have a job you find uncommon to good debut the holiday some people might wonder why I waited that actually we want you to make everyone aware what is the value of a real talent in Huawei and it that's also a gesture of a welcoming real parent minds to join forward just a world or so go to go kosher romantic sometimes in here there's a girl - choco Oh Utica champion took a bigger house Alicia sighs yeah it's a sushi shop she assumed and the girl and her family just there's one story which I shared in the past for the Da Vinci architecture Houston Ascend chips the decision to initiate data project was not a top-down our decision is actually a bottom-up decision that's also a very good example to tell you how and where talent to the people in faraway can play an important role.

The teacher council he is hiding the secret is a $20 how is in the meaning the one Putin - on the couch at since Idaho may podium for Kayla CES 2009 watch o su tienda de currency rupee on legal why pas de yada her charge institutions Idaho sue for a cigar young fun Huawei sanctum Chile will achieve an champion de sica cavalry to countenance IDO regarding culture learning which is our the area Shindo Johnny Giavotella have to do gotta teach our young mayor Celinda take a Indiana jihad will refer to a sign in the Putin cultural change in kinetic Koosman chiba-chiba cited simply on champion angle Russia try to figure Sargento sue Canyon City College at first as a 90-day extension has not had any impact on Huawei.

Just reconsider when the users away like haha wait a tiger sentence silly a Faton 20 some wit notice how a mood shangkun cadence synopsis sale a consider science class of Banja Luka wrongful tie reality and fashion to ho-hoo equate in Santa for way citation Panda 2 shows you know and that your engine for this you consider putting gum tissue holla as you all know this companies cannot continue working with Huawei that's the case I put engines examine but there it will always be other alternative our companies mmm mr. Santa CNT my you couldn't be a the situation here if we look back 10 years ago there was basically no tools available still people can design now to our chipset San Antonio Chang again you can easily book on you know McCulloch the back woman can walk on the many machines and that's the film a or book for you yeah that's good time definitely there would be challenges we have to tackle the efficiency you would not be as high as past and we may need to take on more were close but this is not something or a challenge that is not that is insurmountable but there's some kind intelligent vegetal to secure the girls you constantly looking further in history Intel already came up with the CPUs in nineteen seventies and the companies you mentioned are just now were not even established.

Omen testament the way the tissue destro Inc I took a way to attend to the Jetta and we believe the European Union will be making the right decision as they choose the future technologies for connecting the cars.

However no matter is a cellular V2X or Wi-Fi for Huawei from business perspective is nothing more than providing a module thank you did you find your wife at you - fine we have both 5g solutions as well as Wi-Fi solutions for connecting the cars.

Indonesian jungle so a chance a Samba imagine that that I can't Society Santa Easter weekend yet kikuchi cosas el camión sang is Santa ho when we pass the change again very different power woman Union to desire a beautiful a and you did the parade head miss Ren once said our revenue this year might be 330 billion US dollars - less than we originally planned but it seems at the least it's not gonna be that big in terms of the specific numbers for the whole year we're gonna have announcement of that number often by the end of the year so maybe a us see that number by then.

As for the open sources strategy for the all scenario AI computing framework - Huawei is not a final answer yet. Once we launch as we go open source on MindSpore in Q1 next year we'll also launch our open source strategy together with it.

The teacher some witnesses in the follow it which would receive a sentence Jolene says to Samsung Trombetta vadacurry Tanegashima yes nasi acquire the horse inside I used to teach Indonesia Indonesian super when you Santo Tomas shine the delicate agenda cause you know sometimes really let Sega - sports ago arm the tanker choco yo Tiana Quincy now we can come to order take a mooch energies on couchy Superboy news app doesn't enjoy adding cause it reminds for him take a button that put a hot you the wind he who is and winds first there's no issue regarding acquiring the necessary components to design an interval of Ascend 910 and MindSpore for women the woman something giant a solution that you question yet Sania Sania Sania degrees in here assign Ascend focuses on AI relates the training since on some human condition decision Han so Jung da da da da daaaa there are only two companies working on comparable chipset in the market and I think you all know which those two companies are today one incident yo-yo didn't occur to us guys will see me me your uncle I took unfit what means yield to uncles we part ago the you to associate Ascend 910 is SOC system-on-chip and there are ARM cores our views are in it in Huawei has a perpetual license architectural license of v8 from ARM so um to go simply in the FATA episode employees there for the development of this chipset will not be affected whatsoever.