Lei Gang from Niuaofei Temple
Quantum Bit Report | Public Account QbitAI

What changes are taking place in China's AI?

There is no data more accurate, more intuitive, and more reflective of the wind direction than server data.

Today, IDC released the "2018 China AI Infrastructure Market Data Report", which disclosed the latest developments in the AI ​​server market.

The report includes:

• Market share and sales

• Security and transportation applications are growing rapidly

• Online reasoning is growing rapidly

Details are as follows:

Market share ranking

In 2018, sales in China’s AI infrastructure market increased by 132% year-on-year, with overall sales reaching US$1.318 billion.

Among them, for reference, Inspur, which ranks first, has sales of US$660 million (RMB 4.55 billion), still accounting for more than 50% of the overall share, and has ranked first with more than half of the market share for two consecutive years.

The top five manufacturers are: Inspur (50.1%), Huawei (14.5%), Sugon (10.7%), Dell (6.6%), and H3C (5.7%).

In terms of shipments, the top five manufacturers are: Inspur (47.2%), Huawei (14.6%), Sugon (11.6%), Dell (7.6%), and H3C (6.4%). Inspur shipped 25,551 units in 2018, with a share of 47.2%.

Industry application trends

According to the report, the Internet remains the largest vertical in the GPU server market, but end users in traditional industries are beginning to adopt AI solutions and deploy GPU servers.

Data shows that the top five industries in terms of AI infrastructure scale in 2018 were Internet (63.2%), government (15.1%), services (7.0%), education (5.0%), and finance (2.7%).

Although the Internet is still the largest market, its share has declined for the first time, from 71.6% to 63.2%.

Thanks to the promotion of AI application scenarios such as security and city brain, the share of the government industry has doubled from 7.5% in 2017 to 15.1% in 2018.

The growth rate data also shows the acceleration of AI penetration in industries. The top five industries are transportation (1759%), communications (1720%), public utilities (1214%), health (501%) and government (464%).

The IDC report points out that "traditional industry users are the iceberg hidden under the sea" and that finance, telecommunications, retail, medical care, manufacturing, media, government, transportation, etc. will become the dominant industries of AI in the future.

At present, the main AI technologies used by users in traditional industries are video/image recognition and audio recognition, and their business scenarios include anti-fraud, video surveillance and video analysis, intelligent customer service, precision marketing, medical image recognition, etc.

For traditional industry users, the biggest challenge in applying AI is the lack of technical capabilities in platform, application development, and model and algorithm adjustment.

Online inference acceleration cards are growing rapidly

It is worth noting that acceleration cards used for online reasoning scenarios achieved rapid growth in 2018.

IDC data shows that the sales of M4, P4, and T4 GPUs increased from US$48.1 million in 2017 to US$360 million in 2018, and the market share increased from 8.5% to 27.6%.

Among them, P4 sales increased by 691% year-on-year. At the same time, FPGA sales reached US$13 million in 2018, a sharp increase of 1757.1% year-on-year, and the market share increased 10 times compared with 2017.

This trend shows that China's artificial intelligence has gradually entered the stage of large-scale application, and the process of industrial AI is accelerating.

16-card GPU server is gaining momentum

The report also disclosed changes in accelerator types (GPU/FPGA/others): GPU still leads the market development, accounting for 99% of the market in 2018; FPGA's share increased from 0.1% in 2017 to 1% in 2018.

Among them, GPU servers continued to maintain rapid growth in 2018, with sales increasing by 131.2% year-on-year. They are still the mainstream in the field of AI infrastructure, accounting for 99% of sales.

Nearly 80% of sales come from GPU servers with 4 or more cards, accounting for 80.4%.

The sales of 16-card GPU servers grew rapidly, with sales increasing from US$20.9 million in 2017 to US$263 million in 2018, a year-on-year growth rate of 1161.7%. The proportion of 16-card GPU servers in the overall GPU server market increased rapidly from 3.7% in 2017 to 20.4% in 2018.

The rapid growth trend of 16-card GPU servers indicates that with the explosive growth of data and the increasing complexity of deep neural networks, a single system can provide a higher-density, higher-performance AI infrastructure, which will be more widely used in AI offline training scenarios.