Leveraging Graphcore’s IPU architecture for large scale GNN compute

Machine Learning on large scale graphs presents several unique challenges, due to the sparsity of the connections. Exact computation is often intractable on current accelerators, and algorithmic approximations fall short of modelling interesting aspects like long range dependencies effectively.

We present how Graphcore’s IPU design tackles these challenges, creating the opportunity to accelerate deep GNNs on large graphs. This talk aims at stimulating Data Scientists, Machine Learning Researchers and Engineers to think about different ways to deploy current large scale GNNs and to develop algorithms that exploit the full potential of our new hardware architecture.

Talk+Live Q&A at the Eastern Auditorium in Connected Data World Center

You need an access pass to attend this session: Diversity Access Pass or Full Access Pass apply