In December 2015, Baidu’s driverless car passed a road test starting from the Baidu Tower in Zhongguancun High-tech Park in Beijing, driving onto the G7 Beijing-Xinjiang Expressway, through the Fifth Ring Road and reaching Olympic Forest Park, and back. The driverless technology adopted throughout the course included decelerating, lane switching, overtaking, going on and off ramps and turning around. The car passed each of these testing criteria and reached a top speed of 100 km/h (62 mph).

Image recognition requires hundreds of thousands—in some cases even billions—of learning samples to train a model. A GPU made up of thousands of smaller and more energy-saving cores has therefore become the main force for the application of image recognition training.

Inspur is a leading solution provider of an artificial intelligence application and has a strategic partnership with NVIDIA (the largest vision computing company in the world) in the field of GPU heterogeneous computing.

The NF5568M4 co-processing acceleration server is based on the Intel Haswell E5-2600v3 platform combine with four Nvidia Tesla K40 GPUs. The highest single computing capacity of a single GPU server reaches 17 teraflops.

Presently, in the KITTIKITTI test suite, Baidu has reached a recognition accuracy of 90%—due to a big contribution by Inspur’s GPU co-processing acceleration server NF5568M4.

Baidu plans to set up demonstration regions in 10 cities within China and officially put commercial-purposed driverless cars to use. Large scale production is projected in five years; within 15 years it is estimated that 80% of newly manufactured cars will be equipped with the driverless function.