Data-Intensive Computing Acceleration with Python in Xilinx FPGA

摘要

Data-intensive workloads drive the development of hardware design. Such data intensive services are driven the raising trend of novel machine learning techniques, such as CNN/RNN, over massive chunks of data objects. These services require novel devices with configurable high throughput in I/O (i.e., data-based model training), and uniquely large computation capability (i.e., large number of convolutional operations). In this paper, we present our early work on realizing a python-based Field-Programmable Gate Array (FPGA) system to support such data-intensive services. In our current system, we deploy a light layer of CNN optimization and a mixed hardware setup, including multiple FPGA/GPU nodes, to provide performance acceleration on the run. Our prototype can support popular machine learning platform, such as Caffe, etc. Our initial empirical results show that our system can perfect handling all data-intensive learning services.