Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 14:19
May 19, 2020 14:19
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
August 9, 2020 19:27
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
May 19, 2020 11:43
August 29, 2020 11:48

G3: A Programmable GNN Training System on GPU

What's new

G3 is research prototype that is currently in active development. You are welcome to share ideas and contribute.

The web-based GUI can be found at

VLDB 2020 demo video (with a case study on predicting COVID in Singapore):


G3 is built as a programming framework for Graph Neural Network (GNN) applications that supports graph related operations natively and leverages on the computation power of Graphics Processing Units (GPU). It is named G3, which stands for Graph processing system but tailored for GNN processing on GPUs. The values of G3 can be summarized as follow:

  • G3 thrives to fundamentally enhance the performance of GNN applications by incorporating graph oriented operators at the low level.
  • G3 provides a set of existing basic modules that can be stacked up to build up user-customized GNN architectures.
  • G3 provides flexible C++ based APIs such that developers are able to define and implement their own GNN layers when such layers do not exist in our system.
  • G3 frees the programmers from low level GPU thread management and memory management concerns, which is usually daunting and injects numerous unnoticeable bugs into the system.
  • G3 automatically applies graph-centric optimizations on the GNN architectures during run time.

We have done some preliminary studies to show how such system can be used to build Graph Convolutional Network (GCN) , the state of the art in various GNN models, and how our implementation exceeds the performances of Tensorflow implementation (the original implementation) as well as PyTorch implementation by a significant margin. Details are here.

G3 is built based on Gunrock, a CUDA library for graph-processing designed specifically for the GPU. For more details, see Gunrock's Github repo. Most of the code is reused from Gunrock's code base. G3 added these components to Gunrock to efficiently support GNN training 1) GCN App and 2) GCN Example.

Getting Started


CMake >= 3.10
G++ >= 7.5
CUDA >= 9.0

Quick Install

git clone --recursive
cd G3
mkdir build
cd build
cmake ..
make -j

After that, you can run the example GCN application:

cd bin

How to cite G3

If you use G3 in your paper, please cite our work (full version).

 author = {Liu, Husong and Lu, Shengliang and Chen, Xinyu and He, Bingsheng},
 title = {G3: When Graph Neural Networks Meet Parallel Graph Processing Systems on GPUs},
 journal = {PVLDB (Proceedings of the VLDB Endowment), demo paper},
 year = {2020}

Related work


  • We acknowledge Ee Ter Low, Long Sha, and Karan Sapra for their contributions on developing the demo (with special support under COVID internship program (T1 251RES1824)).
  • This work is supported by a collaborative grant from Microsoft Research Asia.