This market research report was originally published at Tractica's website. It is reprinted here with the permission of Tractica.
Huawei announced its AI strategy at the Huawei Connect 2018 conference in Shanghai. Tractica was invited to the conference with AI being a central theme. In the past, Huawei Connect has been focused on its enterprise cloud services business, with AI being mentioned on the sidelines. This year, Huawei had “Activate Intelligence” as the conference tagline, with more than 25,000 people in attendance filling up Shanghai’s Expo and Convention Center.
Huawei’s Full-Stack AI: Chips, Frameworks, and the Cloud
Huawei’s Chairman and Rotating CEO, Eric Xu, gave the keynote speech announcing the company’s AI strategy covering hardware, software, and services. Huawei’s AI hardware includes chipsets, but it also extends to accelerator cards, AI appliances, and AI servers. Huawei is introducing two new AI chipsets: the Ascend 910 and Ascend 310. The Ascend 910 is focused on data centers with 256 Teraflops (TFLOPs) of compute power (FP16), built on 7 nanometer (nm) process and consuming 350 W of power. The Ascend 310 processor built on 12 nm process is focused on low-power applications (3 to 10 W) like laptops and PCs and provides 8 TFLOPs of compute power (FP16). The Ascend 910 is 2X more powerful than NVIDIA’s V100 solution, which makes Huawei the leading AI chipset company based on performance. The Ascend 310 is available now, while the Ascend 910 will generally be available in 2Q 2019 and, therefore, in terms of what is available currently, NVIDIA still holds the lead.
The Ascend 310 is seen as a versatile processor with multiple combinations of the chipset embedded in several devices under Huawei’s Atlas brand, ranging from accelerator cards, modules, and edge stations to appliances and mobile data centers. Huawei has a common AI architecture across all of its Ascend chipsets called DaVinci. Details of the DaVinci architecture have not been provided, but it includes scalable memory, compute, and on-chip interconnections.
Huawei also has an extended range of Ascend chipsets planned for ultra-low power solutions including Ascend-Nano (1 mW), Ascend-Tiny (10 mW), and Ascend-Lite (1 to 2 W). These are targeted at earphones, Internet of Things (IoT) always-on devices, and smartphones, respectfully. Details about availability for the Nano, Tiny, and Lite were not available, but they are expected to come sometime in 2019. The smartphone-focused Ascend-Lite will not be replacing the Kirin chipset, according to Huawei, but Lite is known to support training on the smartphone with model sizes up to 100 MB and 1 to 10 TOPS of compute capacity.
Huawei has made it clear that it does not intend to offer its chipsets as standalone products and will only offer them as part of an integrated AI solution that includes an appliance, server, card, or other device. Huawei did hint that, as it extends into the edge, it could consider offering its chips separately, although the details are unclear. If that does indeed happen, it would propel Huawei into an entirely new business model, becoming a solutions supplier as well as an independent hardware vendor (like Samsung), which, in the long run, could be work out in its favor as it addresses a larger market.
The hardware chipsets and devices provide the foundation of Huawei’s AI strategy, but the full-stack solution includes a software development framework called MindSpore and a compiler called CANN (Compute Architecture for Neural Networks), which optimizes models for its embedded hardware stack. MindSpore is seen as an alternative to other frameworks like TensorFlow, Caffe, and other AI frameworks, but it can also integrate models from these frameworks. The icing on the AI stack is Huawei’s ModelArts platform, which covers the entire development workflow, from data acquisition, model training, model management, and deployment to adapting to model changes.
Huawei’s ModelArts platform is part of its Cloud Enterprise Intelligence (EI) solution, which offers pre-integrated vertical solutions and hierarchical application programming interfaces (APIs) across application areas. Cloud EI supports vision, language, and decision-making services, and APIs, such as face detection, optical character recognition (OCR), natural language processing (NLP), image recognition, etc. The pre-integrated vertical solutions supported currently include city, internet, home, vehicle, logistics, healthcare, campus, and manufacturing, all of which can be accessed through public, hybrid, and private cloud services.
On the third day of the conference, Huawei provided more details about its cloud strategy, especially around building a developer ecosystem. This includes an AI marketplace that allows developers to reuse pre-built models and data through APIs. The AI marketplace is expected to be open to all developers that wish to access pre-built models and rapidly build new solutions. As a part of its developer go-to-market strategy, Huawei is investing CNY 1 billion in AI talent development with research institutes and universities in China being the prime targets. Huawei’s goal is to have 1 million AI developers using its AI platform in 3 years.
Huawei’s AI Pivot Brings It in Direct Competition with Google
The scale and ambition of Huawei’s AI strategy can be viewed as a pivot from being a traditional telecoms original equipment manufacturer (OEM) to becoming a leading AI solutions vendor. There has been a lot of speculation about Huawei’s AI strategy, but coming out of Huawei Connect 2018, it is fair to say that the scale and ambition of its strategy is bigger than anyone expected.
Huawei’s AI ambitions make sense when viewed through the lens of China’s own AI strategy. Huawei sees itself as the leading player in China when it comes to offering a full-stack AI solution, bigger and bolder than China’s BAT companies (which include Baidu, Alibaba, and Tencent) that have been in the AI spotlight until now. Where BAT excels in software and applications, and overall control of the mobile and web ecosystem, Huawei is using its hardware expertise, largely telecom and information and communications technology (ICT) based, to extend hardware across verticals and grow horizontally into the software and application stack.
While one could argue that Huawei might have been too late for an AI entry, in real terms, the Chinese AI enterprise market has yet to take off. There is adoption across some key sectors like video surveillance, automotive, consumer internet, and financial services, but we are still a few years away from a full-scale inflection point in AI adoption across China. The reason has been the lack of robust enterprise-grade AI platforms. There is a hardware-software stack emerging in specific areas like automotive (Baidu Apollo), smart home (Baidu DuerOS), and robotics (UBTECH), as was in clear evidence at CES 2018, however, a wider enterprise-grade AI stack that extends from hardware and software through to cloud services has been missing.
In fact, the only other full-stack AI solution on offer is from Google, which develops the Tensor Processing Unit (TPU), TensorFlow framework, and Google cloud services. Google has competition from Microsoft and Amazon in terms of their AI cloud services and framework offerings, but Google is the only one that has a hardware and chipset play with its TPU AI processor. While the BAT companies are also developing their own in-house processors, Huawei’s Ascend solution in conjunction with ModelArts, CANN, and MindSpore is China’s answer to Google’s AI strategy. The fact that Huawei has beaten Baidu, Tencent, and Alibaba is significant.
The Challenge Ahead
Huawei’s challenge now is to convince enterprise customers and AI developers that it can deliver on its promises. Huawei was keen to showcase the momentum that it has already built with customers who are starting to trial and deploy some of the AI solutions announced. For the most part, the partners include Chinese companies, but this also includes some big Western companies like Microsoft, which is known to be already using Huawei’s AI solutions in China to power its Bing search engine.
The recent controversial Bloomberg story about Chinese malicious chips being implanted into servers across U.S. companies does not bode well for Huawei or Chinese hardware companies in general if they want to do serious business outside China. Nevertheless, the large presence of business delegations from the Middle East, Africa, and Latin America does tell you that Huawei could be looking at a vast untapped market for AI solutions in developing regions. How Huawei’s AI solutions will be received in Europe is yet to be seen, although its continued strength within mobile and telecom in Europe places it in good stead.
Within China, Huawei will need to look out for the BAT companies, as they plan their next moves in terms of cloud and software, but especially in AI hardware. Alibaba is the cloud services leader with a large enterprise presence and ET Brain AI platform, and it has plans for an AI chip in 2019. Baidu has already announced its hardware foray with the Kunlun chip and show signs of becoming a full-stack AI provider with its Baidu Brain platform and recently announced EZDL stack. Tencent is expanding its AI cloud offerings and has chipset plans. China is seeing a plethora of AI chipset startups looking at both the cloud and edge-based offerings. The year 2019 will see increased momentum for AI, both in terms of hardware and cloud services, with the maturity of enterprise-grade AI going up a notch. As far as the AI opportunity is concerned, we are still at the tip of the iceberg and, therefore, Huawei will need to continue its momentum and plan 2 to 3 years ahead of its rivals if it wants to succeed.
Huawei’s 1 million developer goal is ambitious and, as it will fund Chinese research universities and labs, it will have to convince developers, as well as enterprises, that it can compete effectively with the BAT and Google companies of the world, expanding beyond its traditional area of strength, which is telecommunications.
Aditya Kaul
Research Director, Tractica
