Baidu Accelerates Self-Driving Car Effort with GPU-Powered HPC

May 27, 2016
By: Michael Feldman

Google isn’t the only hyperscale company that is developing autonomous vehicles. Chinese internet giant Baidu is also aggressively pursuing this nascent market, albeit with less public fanfare than its American counterpart. This week though, more about the project’s inner workings was revealed when Inspur announced that its GPU-accelerated servers had been selected by Baidu as a platform for the company’s deep learning image recognition system.

According to the press release, the Inspur gear will be used to train models for image recognition, which will subsequently be used by the driverless cars to detect pedestrians, traffic lights, and other road-related objects:

“Presently, in the KITTI test for common vehicles, Baidu has reached a recognition accuracy of 90% -- thanks to a big contribution by Inspur's GPU co-processing acceleration server NF5568M4, which guarantees greater safety in the operation of driverless cars.”

The NF5568M4 is Inspur’s “Yitian” series, a dual-socket Xeon server that can support up to four accelerators, either NVIDIA GPUs or Intel Xeon Phi. In this case, the Baidu servers are outfitted with two Xeon “Haswell” E5-2600 v3 processors and four Tesla K40 GPUs. Each box delivers over 17 single precision teraflops.

Like Google, Baidu will be demonstrating its self-driving cars in multiple cities. Large-scale production is scheduled to begin within five years.