LightFPGA: Scalable and Automated FPGA Acceleration of LightGBM for Machine Learning Applications

Abstract

FPGAs are used for high speed machine learning inference, and are proving to be much faster and efficient than CPU. LightGBM is a gradient boosting algorithm that uses decision tree-based learners. In this work, we have developed a library named LightFPGA which extracts details from a pre-trained LightGBM model and generates the corresponding Verilog RTL for FPGA implementation. Since the whole process of code generation is automated, the design is scalable to the LightGBM model trained for any given dataset. Further, the library performs testing and accuracy verification of the implementation by generating testbench. Our results show that using LightFPGA, around 100–400× improvement in latency as compared to CPU can be achieved without any reduction in inference accuracy. Further, it has been observed in the tests performed, that the FPGA implementation of LightGBM offers around 7–8 folds of power reduction, as compared to CPU.