Alibaba unveils its own AI chip for cloud-computing services

Alibaba used its Aspara Conference, held in Hangzhou, China, to reveal the cloud AI accelerator chip – its first such self-developed device – with Jeff Zhang, Alibaba CTO, showing the crowd the new chip.

Called Hanguang 800, the 12nm chip has 17 billion transistors. Alibaba hinted at the new chip's performance by sharing a few benchmark figures, using the popular convolutional neural network ResNet-50. According to Alibaba, the Hanguang 800 chip running the ResNet-50 can process 78,563 images per second (IPS) and operates at a peak efficiency of 500 IPS per watt.

Zhang claimed that this meant Alibaba’s silicon was 15 times and 46 times more powerful than Nvidia’s T4 GPU and its older P4 model respectively.

Although it is not yet in mass production, the AI chip is currently being used by Alibaba to power product search, automatic translation and personalised recommendations on the e-commerce giant’s suite of public-facing websites.

“The launch of Hanguang 800 is an important step in our pursuit of next-generation technologies, boosting computing capabilities that will drive both our current and emerging businesses while improving energy-efficiency,” Zhang said in a statement.

The chip was developed by DAMO Academy, a research institute Alibaba launched in late 2017, and T-Head, the company’s specialised semiconductor division.

Alibaba’s foray into the chip sector comes amid efforts by Beijing to promote China’s semiconductor industry and reduce the country’s reliance on foreign imports of core technologies.

Alibaba has no immediate plans to sell the chip as a standalone commercial product, a company spokesperson later said, although this could change in time. However, the new silicon is expected to be available on the company's cloud servers in due course for customers to rent.

Alibaba also released its first core processor IP in July based on RISC-V open-source chip architecture. RISC-V gives firms a potential alternative to the dominant architecture of Britain’s Arm Holdings. Arm, a unit of Japan’s SoftBank Group Corp, charges licensing fees for its use.

Rivals such as Alphabet Inc (parent company of Google) and Facebook are also known to be developing their own custom chips, in order to improve the performance of specialised AI tasks at their data centres.

Alibaba has become more intently focused on producing its own chip hardware since its ex-CEO Jack Ma said that China should stop relying so much on US silicon imports. Ma set up Alibaba's new semiconductor division, T-Head (aka 'Pingtouge', meaning 'honeybadger' in Chinese), in order to make the company more self-sufficient. The ongoing trade tariff war between the US and China has galvanised these efforts, with technology suppliers in both countries particularly negatively affected by the tariffs.

In cloud-computing terms, Alibaba towers over its rivals in China, commanding 47 per cent of the market for cloud infrastructure services in the first quarter of 2019, according to research firm Canalys.