Bridging the Gap: Accelerating Random Forests on FPGAs with High-Bandwidth Memory
Icaro Moreira, Lucas Bragança, Olavo Silva, Alysson Silva, Ricardo Ferreira, José Augusto M. Nacif · 2025
In memory-bound problems, Field Programmable Gate Arrays (FPGAs) have traditionally underperformed compared to Graphics Processing Units (GPUs) due to their lower memory bandwidth. However, high bandwidth memory (HBM) in FPGAs has significantly enhanced their performance, achieving bandwidths up to 425 GB/s. Additionally, FPGAs offer the advantage of customizable accelerators for domain-specific tasks, potentially outperforming general-purpose GPU architectures. This work focuses on accelerating random forest algorithms on FPGAs, leveraging their customization capabilities to manage control flow structures such as decision branches efficiently. Despite these advancements, FPGAs remain challenging to program, requiring a deep understanding of hardware design. We propose a new hardware generator that integrates necessary tools into a cohesive workflow to address this, simplifying FPGA development. We validated the design on a Xilinx Alveo FPGA, using 32 HBM channels and reaching a performance of 8 billion samples per second. This work offers a practical solution for memory-bound machine learning tasks in high-performance computing environments.