Reproduce “Efficient Recurrent Neural Networks using Structured Matrices in FPGAs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhe Li, Shuo Wang, Caiwen Ding, Qinru Qiu, Yanzhi Wang, Yun Liang

Researchers from Syracuse and Peking Universities developed an FPGA-based accelerator for recurrent neural networks that employs block-circulant matrices and Fast Fourier Transforms. This method achieves up to 24.5x faster inference speed and 35.7x greater energy efficiency on FPGAs compared to state-of-the-art pruning techniques, while maintaining negligible accuracy degradation on the TIMIT speech dataset.

Key insights

  • Structured compression using block-circulant matrices provides superior hardware regularity and higher compression ratios compared to unstructured pruning, enabling more efficient memory access and parallel processing on FPGAs.
  • Integrating FFT into the forward propagation of block-circulant matrices offers a computationally efficient alternative to direct dense matrix multiplication, leading to significant acceleration.
  • Training models directly with block-circulant constraints simplifies the compression pipeline by eliminating the iterative pruning and retraining steps required by traditional sparse methods.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.