Before starting, what kinds of scales can I run GAT on?

Questions
#1

Hi All,

Before diving in, I wanted to investigate on what kinds of scales can I run the GAT implementation on?
I have a graph of 10-20M nodes, each with 2000~ features.
Should that be feasible?

Thanks!

#2

As long as you won’t be running full batch training, and sampling the neighbourhood for message passing if the graph is dense, it should work fine.
But the training times vary heavily with your batch size and CUDA memory, you can check GraphSAINT sampler that is good for the scalability: dgl.dataloading.SAINTSampler

1 Like
#3

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.