TC-GNN - Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs

Explore a groundbreaking conference talk on accelerating Graph Neural Networks (GNNs) using GPU Tensor Core Units (TCUs). Discover TC-GNN, the first GNN acceleration framework that reconciles sparse GNN computation with high-performance dense TCUs. Learn about the novel sparse graph translation technique and effective CUDA core and TCU collaboration design that enables full utilization of GPU resources. Understand how this innovative approach achieves an average 1.70× speedup over the state-of-the-art DGL framework across various models and datasets. Gain insights into the challenges of GNN performance due to sparse and irregular graph-based operations, and how TC-GNN addresses these issues. Presented by researchers from the University of California, Santa Barbara, this 18-minute talk from USENIX ATC '23 offers valuable knowledge for those interested in graph-based machine learning and GPU optimization techniques.