揭秘华为AI计算变阵，他们要卖“挖掘机” | 专访昇腾计算总裁许映童

乾明 发自 凹非寺
量子位 出品 | 公众号 QbitAI

过去这大半年，华为昇腾AI计算业务，已经悄然完成了一场变阵：

It is no longer just a R&D team, the marketing and sales team is in place, and it has become a business organization covering all aspects of product development, sales, and service.

Correspondingly, the focus of the entire business has also been adjusted accordingly: it is no longer limited to technology research and development, but is also accelerating the construction of an ecosystem to promote the implementation of technology in the industry.

现在，华为昇腾AI计算团队已经超过2000人，全处于紧锣密鼓的攻关状态，70%的工作量在“软件”打造上，为各个行业打造基础软件平台，提供工具，输出行业SDK。

Why did Huawei make such adjustments?

华为昇腾总裁许映童接受量子位采访，解读了昇腾计算变阵背后的驱动力，以及他对整个AI计算产业发展的思考。

Xu Yingtong is a Huawei "veteran".

He joined Huawei in 1999 and served as Wireless GPRS PCU PDT Manager, Director of Wireless Base Station Software Development Department, and President of Smart Photovoltaic Business. He created wireless software platforms, photovoltaic inverters and other leading competitive products.

He has extensive experience in the development and application of new technologies. If you are also concerned about the development and changes of the AI ​​industry, his latest sharing is not to be missed.

Behind Huawei's AI computing transformation: strategic execution and customer needs

昇腾AI计算业务变阵，起于2019年9月举办的华为全联接大会结束后。

而在此一年前（2018年），华为刚刚正式对外披露了全场景AI战略。同年11月份的世界互联网大会上，发布了昇腾系列首款芯片昇腾310。

在这两年中，华为发布了训练芯片昇腾910，并围绕昇腾打造了完整的软硬件一体化的解决方案。

现在，华为昇腾计算已经转化成了云端边协同的“生产力”，在新的业务组织架构下，应用到了互联网、运营商、金融、能源、制造等多个行业中。

比如华为与意大利AGS公司等合作伙伴，打造出的基于人工智能的医疗影像辅助诊断系统。在昇腾AI算力加持下，医生和病人获取精准的医疗影像诊断报告的速度大大加快：

It has been reduced from 10-15 minutes to 10-15 seconds. As the global anti-epidemic situation remains severe, such assistance also reflects the inclusiveness of AI.

在华为内部，昇腾AI计算也已经应用到了各个方面。

For example, at Huawei's Songshan Lake manufacturing center, Huawei combined the AI ​​accelerator card Atlas 300 with an industrial computer to create machine vision intelligent quality inspection, which can complete tasks such as label defect detection and missing screws.

Xu Yingtong said that this directly increased the detection efficiency from 90% to greater than 99.9%, and shortened the time from the original 180 seconds to seconds.

Such a fast strategic advancement is truly amazing.

But in Xu Yingtong's view, this is just the tip of the iceberg. Such a speed is due to Huawei's heavy investment in AI technology research and development, which has exceeded US$1 billion so far; on the other hand, it is inseparable from Huawei's technology research and development strategy.

He said that Huawei's technology research and development usually adopts a "three-generation strategy", which are:

In the pre-research generation, key technologies are pre-researched 5 to 10 years in advance; in the R&D generation, technology research and development is carried out around products that will be launched in the next two to three years; and then there is the launch generation that everyone can see.

华为昇腾计算业务，目前正处于研发一代向上市一代的转变中。

许映童说，昇腾AI相关的技术经过提前多年的研发与孵化，已经通过验证，得到了市场认可，处于较为成熟的状态。

At the same time, the demands for practical applications of AI technology from external customers and Huawei itself have gradually increased. Under the previous R&D organizational structure, it was difficult to respond quickly and mobilize resources, so organizational structure adjustments became inevitable.

He said that under the business organization model, the group will also give the team more authorization, and the team will have more decision-making power over business, market, product planning, and customer selection, so that they can quickly adjust according to market and customer needs and the competitive status of the industry.

Xu Yingtong is quite impressed by the current explosive demand for AI computing in all walks of life.

"The market and customer demand far exceeded our expectations," he said. "In contrast to other industries that first incubate products and then promote them, we are being pushed by our customers and our development speed is not fast enough."

尽管昇腾AI计算已经构建起了完善的商业组织，形成了完善的落地方案，并在不同的场景中得到了应用，发挥出了效率。

But business goals are not Huawei's top priority in this field. In contrast, ecology is the first priority.

Huawei plans to sell "excavators" after three years of commercial goals

Xu Yingtong said that AI computing is an ecological industry. Huawei has no real competitors and everyone is a pioneer and forerunner of the industry.

The core competitiveness lies in the group of developers who understand AI. Only they can unleash the computing power of AI and truly apply AI to the industry.

But they are the most scarce resource in the entire industry. There are only more than 30,000 AI experts in the world, and the talent supply-demand ratio is only 1%.

培育更多的人才繁荣生态，是华为昇腾AI计算业务的核心任务，他们计划在未来三年发展100万名开发者。

目前，华为昇腾已经面向高校、开发人员、初创企业等推出了合作伙伴计划，并提供从学习到产品构建，再到产品上市、销售的全方位激励机制。

In Xu Yingtong's view, the ecosystem is related to the building of long-term competitiveness, while the business goals will be focused on three years later.

他说，在构建生态的同时，华为也会进一步完善昇腾产业相关的技术和平台，打造“用得起、用得好、用得放心”的AI计算产品。

Xu Yingtong said that when AI enters various industries, various decision-making processes will emerge due to the wide range of application scenarios, and it will be almost impossible to create highly adaptable products.

Huawei will abstract an SDK for a certain industry from the perspective of foundation and platform, build it into a platform or infrastructure, and make it more complete and mature, forming an integrated solution of end, edge and cloud to promote the implementation of AI in thousands of industries.

最后，该如何看待华为昇腾AI计算的变阵？从整个行业近几年的发展来看，不难找到答案。

What is computing power? The productivity engine under the AI ​​revolution determines the speed and height of this historic revolution and is the key to productivity in the digital economy era.

With the development of technology, the maturity of AI algorithms, the density and quality of AI computing power are improving, the threshold for obtaining it is gradually lowered, and the demand for AI applications in various industries is beginning to be released.

Especially with the arrival of the new infrastructure in 2020, these demands have been magnified exponentially. Xu Yingtong said that this is the most concerned issue for the entire industry at present.

However, if we want to make more effective use of computing power and apply it to scenarios in all walks of life, there are still shackles on the implementation of AI, such as the grasp of demand, the combat effectiveness of the organization, the speed of response to the market, etc.

Why did many technology companies invite top academic giants to join them in the early years to build corporate research institutes, but in the end, nothing came of it and the talents left? It is hard to say that there is no influencing factor in this aspect.

Huawei, which has gone through several technology cycles, naturally sees things clearly and is usually one of the companies with the strongest action and fastest execution.

Looking to the future, there are still ample opportunities for development.

Xu Yingtong shared a set of data: the overall penetration rate of B2C applications containing AI in the Chinese market is 10%, while the overall penetration rate of AI in the industry is only 4%.

With the continuous breakthroughs in AI computing power and algorithms, the number of AI talents will gradually increase, which will obviously generate huge commercial value.

In his opinion, facing such a "rich mine", Huawei is definitely not just a water seller.

“We sell excavators,” he said.