Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A survey on FPGA-based accelerator for ML models

Feng Yan, Andreas Koch, Oliver Sinnen

TL;DR

This survey addresses the landscape of FPGA-based accelerators for ML by analyzing 287 papers from four premier FPGA venues over the last six years. It reveals a strong focus on inference (about 81%) over training (about 13%), with CNNs dominating CNN-focused acceleration and growing attention to GNNs and Transformer-style attention. The authors provide a taxonomy across four categories, detailed model-specific accelerators, and a set of optimization strategies (quantization, pruning, mixed precision, memory hierarchy, and matrix-operation specialization) that shape current designs. The study highlights key trends, including maturation of CNN accelerators, emerging attention and GNN work, and ongoing challenges in training acceleration, memory bandwidth, and hardware–software co-design, offering practical guidance for future FPGA AI hardware development.

Abstract

This paper thoroughly surveys machine learning (ML) algorithms acceleration in hardware accelerators, focusing on Field-Programmable Gate Arrays (FPGAs). It reviews 287 out of 1138 papers from the past six years, sourced from four top FPGA conferences. Such selection underscores the increasing integration of ML and FPGA technologies and their mutual importance in technological advancement. Research clearly emphasises inference acceleration (81\%) compared to training acceleration (13\%). Additionally, the findings reveals that CNN dominates current FPGA acceleration research while emerging models like GNN show obvious growth trends. The categorization of the FPGA research papers reveals a wide range of topics, demonstrating the growing relevance of ML in FPGA research. This comprehensive analysis provides valuable insights into the current trends and future directions of FPGA research in the context of ML applications.

A survey on FPGA-based accelerator for ML models

TL;DR

This survey addresses the landscape of FPGA-based accelerators for ML by analyzing 287 papers from four premier FPGA venues over the last six years. It reveals a strong focus on inference (about 81%) over training (about 13%), with CNNs dominating CNN-focused acceleration and growing attention to GNNs and Transformer-style attention. The authors provide a taxonomy across four categories, detailed model-specific accelerators, and a set of optimization strategies (quantization, pruning, mixed precision, memory hierarchy, and matrix-operation specialization) that shape current designs. The study highlights key trends, including maturation of CNN accelerators, emerging attention and GNN work, and ongoing challenges in training acceleration, memory bandwidth, and hardware–software co-design, offering practical guidance for future FPGA AI hardware development.

Abstract

This paper thoroughly surveys machine learning (ML) algorithms acceleration in hardware accelerators, focusing on Field-Programmable Gate Arrays (FPGAs). It reviews 287 out of 1138 papers from the past six years, sourced from four top FPGA conferences. Such selection underscores the increasing integration of ML and FPGA technologies and their mutual importance in technological advancement. Research clearly emphasises inference acceleration (81\%) compared to training acceleration (13\%). Additionally, the findings reveals that CNN dominates current FPGA acceleration research while emerging models like GNN show obvious growth trends. The categorization of the FPGA research papers reveals a wide range of topics, demonstrating the growing relevance of ML in FPGA research. This comprehensive analysis provides valuable insights into the current trends and future directions of FPGA research in the context of ML applications.

Paper Structure

This paper contains 23 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Computing process
  3. Inference vs Training
  4. Reasons for dominance of Inference
  5. Low Latency Requirements
  6. Efficiency Considerations
  7. Challenges and Potential of Training Acceleration Research
  8. Data Processing Complexity
  9. Algorithm Complexity
  10. The Fundamental Role of Matrix Operations
  11. Accelerators for different models
  12. Dominance of CNN
  13. Quantization
  14. Pruning
  15. Growing period (2018-2019)
  16. ...and 8 more sections

Figures (5)

  • Figure 1: Distribution of FPGA accelerator directions
  • Figure 2: ML Related in past 6 years
  • Figure 3: Computation acceleration proportion
  • Figure 4: Distribution of Model Types in Inference and Training Paper
  • Figure 5: Surveyed paper numbers on typical ML models by year