Confident leading in artificial intelligence but turned out not to have components from the US, China had to give up

In March 2018, when a group of Chinese artificial intelligence (AI) leaders were about to fly to Silicon Valley to attend Nvidia’s annual GPU Technology Conference, they were hosted by instructing and lowering relations with US technology firms in the context of increasing US-China trade tensions at that time.

” We were asked to remove the press releases indicating our close relationship with Nvidia, and we did as requested .” A representative of a leading Chinese AI company told SCMP, but asked to remain anonymous about the sensitivity of the issue. ” But we all felt a chill at the time, and I knew that many Chinese AI companies started to prepare for Plan B when they returned to China after that event .”

Both China and the United States are vying for superpower positions on the AI ​​playing field, with several Chinese companies gradually expanding beyond the domestic market.

Heavily affected by the US ban on Chinese technology companies

China’s emerging AI champions, however, are facing a tough test when eight of them were blacklisted last week by the US, and prevented them from acquiring U.S. technology.

These companies include SenseTime’s face recognition startups, Megvii Technology and Yitu Technology, Hangzhou Hikvision Digital Technology and Dahua Technology video surveillance equipment production companies, AI iFlyTek, Xiamen Meiya Pico Information Co. and Yixin Science and Technology Co.

” The business portfolio of these 8 AI companies is primarily about software and solutions – so the impact of the ban is relatively limited .” Charlie Dai, chief analyst at Forrester Research, said. ” However, it is not easy to find alternative supplies for the AI ​​hardware needed, such as sensors, GPUs, FPGA chips and chipset design software .”

Megvii face recognition software.

FPGA is the abbreviation for reprogrammed array chips, meaning that integrated circuits can be reprogrammed after production.

On Monday, CEO and cofounder of AI startup Megvii said the ban was a “challenge” and it would have an impact on their server supply – as well as the IPO plan on the exchange. Hong Kong stock.

” Its specific impact is that we cannot directly buy certain products due to the US export ban, for example x86 server processors and GPUs manufactured by this country .” Yin Qi, co-founder and CEO of the company, said in an internal meeting with employees.

Anyway, Yin said Megvii was ” well equipped for this battle ” and began diversifying its supply in May, when telecommunications equipment maker Huawei Technologies was first included. black list of america.

The x86 server processors are a family of instruction architectures originally developed by Intel on 8086 processors. Despite a series of Chinese brands such as Inspur and Sugon that have announced server processors They still have to rely largely on patent licensing from US semiconductor giants like Intel or AMD.

In the case of GPUs, Chinese vendors cannot even compete with AMD or Nvidia in the commercial product market.

The image shows the use of face recognition technology at the Artificial Intelligence Conference in Shanghai in August.

According to a report from CB Insight, China is currently the country with the most AI startups worth over US $ billion. Six of the world’s top 11 unicorns in AI come from China, with SenseTime currently leading the pack at a valuation of $ 4.5 billion.

Although Chinese AI startups raised US $ 4.9 billion in 2017, surpassing US rivals with a raised capital of only US $ 4.4 billion, many analysts said many Chinese companies is still heavily dependent on US-made AI servers – devices necessary for the development of deep learning algorithms and techniques.

As the US closes the door to China, what can Chinese tech firms do?

As noted by research firm Jefferies last week, China’s leading surveillance camera supplier, Hikvision Digital Technology, could find an alternative supplier for U.S. components such as cameras and storage devices. from Chinese companies and Japanese suppliers.

However, Hikvision still relies heavily on Intel and Nvidia for CPUs and GPUs to operate AI servers.

” Nvidia’s GPU outperforms other solutions in deep learning, thanks to its CUDA toolkit and libraries. ” Report from analyst Jeff Wu and Lydia Lin of Jefferies Research said. CUDA offers high performance as well as built-in AI functions.

Hikvision camera is used in surveillance activities.

Kuang Kaiming, an AI engineer from Diannei BioTechnology in Shanghai, said that CPU and GPU are essential for deep learning technology, which needs to process huge chunks of training data on hardware.

” They (CPUs and GPUs) are like roads. No matter what drive you drive, they still need roads to go .” Kuang said.

The lack of competitiveness in the core areas of strategic technology is one reason why Chinese President Xi Jinping has always called for self-reliance and resilience – it’s not just a business issue, but also to consolidate military strength and economic security.

The problem became even more urgent when the US government brought ZTE, a Chinese state-owned company, to the brink of bankruptcy in violation of the US embargo. Then there’s the blacklisting of Huawei in May. So far, about 180 Chinese entities have been blacklisted by the United States. For Russia it is from 310 to 320 entities.

China is closing the gap on AI chips with the US

While China also has some alternatives to products from Intel and Nvidia, they do not have the same quality as American companies.

” Leading Chinese technology firms like Alibaba, Huawei and Baidu are making strategic investments in AI hardware, some for training and others for reference research .” Forrester analyst Dai said. ” However, the feasibility of replacing chipsets depends on many other factors, such as product completion rate, server compatibility, software interaction and workload capabilities .”

Xuantie 910 processor – Alibaba’s AI processor.

A number of Chinese companies are making positive strides with AI chips, so they can handle a variety of applications, including IoT devices and self-driving cars.

Alibaba Group Holding, the e-commerce giant and China’s leader in cloud computing, last year established a chip subsidiary called T-head (or Pingtouge in Chinese, nicknamed). of a badger) to make their own reference AI chips.

In July, they unveiled the first Xuantie 910 processor – which can be used in areas such as 5G mobile networks, AI and self-driving car controls.

At the end of August, Huawei revealed its high-end AI chip – Ascend 910 for servers – with the claim it is the ” world’s most powerful AI chip ” aimed at tasks such as training AI models. twice the performance of Nvidia’s Tesla V100.

The Shenzhen-based company earlier revealed a new AI computing system, Atlas 900 with thousands of Ascend processors inside. Huawei said it has set a new world record for AI training performance.

According to Jefferies’ report, Huawei’s Application-specific Integrated Circuit (ASIC), along with local AI chips from companies like Horizon Robotics and DeePhi Technology, not only bring AI solutions. advanced, but also lower the cost for AI chips from 30% to 40%.

The Tianjic processor is tested on a self-driving bike.

In August, a team of researchers from Tsinghua University in Beijing created a Tianjic chip, a hybrid processor, with architecture that simulates neurons but still be able to run deep learning algorithms like common chip. These chips have more efficient power consumption than conventional integrated circuits because they only work when required, in the same way that neurons in the brain “stimulate.”

Research organization Center for Data Innovation sees these signs that China has begun to close the gap with the US, at least in the field of AI chips. According to its research, China has only one company in the top 15 companies in semiconductor sales in 2019, while Europe has 2 and the US has 6 companies.

But based on the number of AI chip design firms in 2019, the gap has narrowed – the U.S. has 55 companies, followed by China with 26 while Europe ranks third with 12 companies.

According to Jefferies’ report, even in the worst case, when Chinese companies like Hikvision are banned from accessing any GPU again, they can still ” train their algorithms through Hubs.” Chinese Supercomputer “.

However, relying on the cloud power of third parties to train AI can cause security issues, when AI companies will need to share data with that third party. According to Xia Yin, an AI scientist in Silicon Valley, this is against privacy agreements with users.

Refer to AbacusNews