Reproduce “RTop-K: Ultra-Fast Row-Wise Top-K Selection for Neural Network Acceleration on GPUs” 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

Xi Xie, Yuebo Luo, Hongwu Peng, Caiwen Ding

RTop-K is an algorithm and GPU kernel designed for ultra-fast row-wise top-k selection, specifically accelerating Graph Neural Network (GNN) training by optimizing a bottleneck operation introduced by MaxK-GNN. It achieves up to an 11.49x speed-up over PyTorch's `torch.topk` and contributes to 11.97% to 33.29% overall GNN training speed-ups, with minimal impact on model accuracy.

Key insights

  • Neural networks, particularly GNNs, exhibit robustness to controlled approximations in top-k selection, allowing for performance gains through techniques like early stopping without compromising accuracy.
  • Optimizing GPU kernel design with warp-level primitives and efficient shared memory usage is crucial for achieving high throughput in scenarios involving a large batch of small, independent computations.
  • Tailored algorithmic solutions are necessary to address specific computational patterns and bottlenecks arising from new neural network architectures, distinguishing them from general-purpose algorithms.

Summary by alphaXiv.

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