FPGAs Accelerate Machine Learning at Baidu

Print Friendly, PDF & Email

xilinxXilinx has announced that Baidu, a Chinese language Internet search provider, is utilizing Xilinx FPGAs to accelerate machine learning applications in its datacenters in China.

The two companies are collaborating to expand the volume of FPGA-based deployments to accelerate its machine learning platforms. The rapid growth of emerging machine learning applications is increasing the need for computation in the data centre. To overcome this new challenge, datacenters operators are turning to application accelerators to keep up with the demands for greater throughput at low latency while retaining acceptable power levels.

“Acceleration is essential to keep up with the rapidly increasing data centre workloads that support our growth,” said Yang Liu, executive director at Baidu.

“Xilinx FPGAs are assisting greatly with this critical task and can provide significant value in the design of autonomous vehicles,” added Junwei Bao, director in Baidu’s Autonomous Driving Unit.

FPGAs are not a new technology in datacenters but they can be difficult to optimize as the hardware of an FPGA must be tuned for each application – which can be time-consuming for datacenters with varied workloads. However, as machine learning applications rely on processing huge data sets, these monolithic workloads are well suited to FPGA acceleration.

Xilinx FPGAs deliver the power efficiency that makes accelerators practical to deploy throughout the data centre. In addition, the company claims that FPGAs can deliver a 10-20X performance/watt improvement compared to other accelerator technologies.

The Baidu-optimized FPGA platforms are tuned for machine learning applications such as image and speech recognition. But the platform will also be leveraged in Baidu’s initiative to develop commercially viable autonomous cars. Baidu is using this technology to provide a centralized pool of accelerators that can be rapidly configured for the most demanding machine learning workloads.

The momentum for FPGA-based acceleration continues as shown by this significant implementation with Baidu,” said Victor Peng, executive vice president and GM of the Programmable Products Group at Xilinx. “We celebrate Baidu’s innovation, expertise and creativity in bringing advanced applications to market.”

This story appears here as part of a cross-publishing agreement with Scientific Computing World.

Sign up for our insideHPC Newsletter