Xilinx, the world’s leading designer and supplier of programmable logic devices, today announced its acquisition of DeePhi Tech — a Beijing-based chip unicorn with a focus on machine learning, specializing in deep compression, pruning, and system-level optimization for neural networks.
Xilinx is best known for inventing Field Programmable Gate Arrays (FPGAs), a type of processor particularly powerful in small-scale but intensive data access. The FPGA chip allows users to program the circuit path through its tiny logic block to handle any kind of digital function. With advantages in low-latency streaming and data-intensive tasks, FPGAs are well suited to cloud computing industries. Microsoft has been deploying FPGAs in its Azure servers for many years. AI researchers believe FPGAs have huge potential in the area of training, where Nvidia’s GPU is currently a dominant player.
Xilinx has been keeping an eye on DeePhi since it was founded in March 2016. DeePhi provides end-to-end solutions using a deep-learning processing unit (DPU) platform and Deep Compression, a technique that aims to compress neural nets by an order of magnitude without losing prediction accuracy. Deep Compression was introduced by DeePhi Chief Scientist Song Han in 2015. The company also specializes in pruning and system-level optimization for neural networks. DeePhi’s founders and employees are mostly from Tsinghua University.
DeePhi has attracted approximately US$100 million in three financing rounds from investors including GSR Ventures Capital, Samsung Venture Capital, Xilinx, MediaTek, and Ant Financial. It is China’s upcoming unicorn in AI hardware, alongside Cambricon Technologies and Horizon Robotics.
On May 22nd, 2017, Xilinx announced an investment in DeePhi after the company won the Best Paper award at the FPGA 2017 with ESE: Efficient Speech Recognition Engine with Sparse LSTM. The DeePhi engine performed 43 times better than CPUs with 40 times the performance per unit of power. Performance was three times that of GPUs with 11 times less power consumption.
DeePhi’s technologies have already been deployed on Xilinx FPGAs, as one of the company’s earliest product solutions.
Financial terms of the transaction were not disclosed. After the acquisition, DeePhi’s 200 employees will continue to operate out of the same Beijing offices.
“We are excited to continue our strong partnership with Xilinx and work even more closely to deliver leading machine learning solutions to our customers in China and around the world,” said DeePhi CEO Song Yao in a statement released today.
“Xilinx is accompanying DeePhi Tech along its journey to explore the potential of machine learning and is supporting our innovation as one of our early investors. We look forward to continuing our joint efforts with Xilinx to bring our solutions to the next level in performance,” said DeePhi CTO Yi Shan.
Journalist: Synced Global Editorial | Editor: Michael Sarazen