Alibaba Group’s cloud computing arm launched a one-stop platform on Tuesday to accelerate artificial intelligence development by corporate clients.
The Apsara Intelligence Computing Platform shortens the training time for AI algorithms in data-intensive fields. Chinese electric vehicle manufacturer XPeng is one of the first companies to deploy it.
The platform can boost AI training efficiency by 11 times and raise algorithms’ inference efficiency — the process by which models make predictions — by six times, according to Alibaba Cloud.
“As the demand for self-driving algorithm training increases, we need robust local and cloud-based computing power to provide support,” said XPeng Chairman and CEO He Xiaopeng.
Training a deep learning model is resource-intensive.
The amount of computational power required to train the world’s largest AI models has almost doubled every three months since 2012, according to San Francisco-based research group Open AI.
But companies are finding ways to increase efficiency. Alibaba Cloud’s latest computing platform will increase the training speed of the XPeng autonomous driving model by nearly 170 times, from seven days to an hour.
In addition to cutting down on energy use, corporate clients on the Intelligence platform will receive access to database services, machine learning platforms and pre-trained AI models.
The Intelligence Computing Platform will enable Xpeng to develop safe vehicles capable of operating autonomously under various conditions.
As it stands, the automaker’s highway driving feature allows its vehicles to change lanes automatically, adjust speed, overtake cars and enter or exit highways.
But to function on urban streets and under unpredictable road conditions, Xpeng needs more computing power to train its algorithm.
“Companies in self-driving research and development have to face the challenge of balancing the need for computing power, maintaining cost control, and seeking tech development,” He said.
On Tuesday, Alibaba Cloud said that a new computing center in Inner Mongolia in northern China had launched to provide infrastructural support for Xpeng’s AI training.
All intelligent vehicle makers will have to self-develop their smart driving features in the future, said He.
Training is a crucial stage in developing an AI algorithm, and this energy-intensive process is getting a make-over, thanks to Alibaba Cloud.
The cloud computing unit uses the Intelligence Computing Platform to train its mega AI model, M6, the world’s first 10-trillion-parameter pre-training model.
M6 can complete the training in just 10 days thanks to the platform’s 512 graphics processing units (GPU), which power data processing and graphics rendering.
“The increase in scale in data and AI models calls for more efficient, less costly and greener intelligent computing power,” said Cai Yinghua, President of Global Sales at Alibaba Cloud Intelligence.
Alibaba Cloud has also tapped the Intelligence Computing Platform to develop its image search AI model for reverse searches on the group’s e-commerce platforms.
This cuts the time it takes to train an algorithm to parse through one billion images from 2.5 months to eight hours.