TC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
summary
TC-GNN implements a GPU pipeline that turns irregular GNN sparse ops into dense tiles processed on tensor cores, combining sparse graph translation with a CUDA+TCU collaboration runtime and PyTorch integration. Evaluation across small to highly irregular real graphs shows an average 1.70× speedup over DGL, operator gains of ~1.5× for SpMM and ~6.9× for SDDMM, lower memory overhead than pure densification, and stable end-to-end training and inference improvements.
AI-written summary · about
chapters (9)
- 0:00Introduction to GNN
- 0:57GNN computation flow
- 2:06GPU acceleration challenges
- 3:54Design approach analysis
- 6:15The TCGN framework
- 8:04Sparse graph translation
- 10:27Data flow and execution
- 12:57Performance evaluation
- 16:40Future work and conclusion
related talks
19mOptimizations and abstractions for sparse machine learning
19mNeuGraph: Parallel Deep Neural Network Computation on Large Graphs
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
13mGNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
16mPreview: Graph NNs
30mUnderstanding GPUs
15mIdentifying and Analyzing Pitfalls in GNN Systems