Xilinx this week announced that it had taken over DeepPhi Technology, a machine learning startup from China. DEEPhi has been using Xilinx FPGA for its ML projects since its inception in 2016 and is therefore seen as a good fit for Xilinx to expand further into machine learning.

DEEPhi’s key product is Xilinx FPGA-based Aristotle architecture that is used to compute convolutional neural networks (CNN). The product is now used for video and image recognition tasks on surveillance cameras and NVR/DVR solutions, but the architecture itself is flexible and scalable for everything from smartphones to servers. For example, right now DeepPhi’s NVR/DVR solution can analyze up to nine channels of 1080p videos in real time and can model for over 30 human faces in a single frame. Obviously, the platform may support other applications too and this might be a reason why Xilinx decided to acquire the company.

Meanwhile, the key features of the DEEPhi Aristotle are the DECENT (DEep ComprEssioN Tool) neural network compression tool used to compress model sizes from 5x to 50x without loss of accuracy as well as DNNC Neural Network Compiler optimized for DEEPhi's instructions.

Since DEEPhi is a privately owned company, terms and conditions of the deal are not disclosed beyond the fact that the company will continue to work with its clients going forward.

Related Reading

Source: Xilinx

Comments Locked


View All Comments

  • mgl888 - Thursday, July 19, 2018 - link

    Pedantic correction: it's DeePhi not DeepPhi
  • mode_13h - Thursday, July 19, 2018 - link

    Had to LOL at "High Speed Data Tube". First time I've ever heard of a "Data Tube". I assume it should probably be "Data Pipe".

    ...not to imply that this is AT's error, or that they should correct PR materials provided to them.

Log in

Don't have an account? Sign up now