Speaker
Andrzej Nowak
(CERN)
Description
Affordable and efficient accelerators are becoming a standard data-center offering. However, programming these devices requires a shift to a heterogeneous approach, discussed here, where data access between nodes can be expensive and floating point answers obtained on different hardware can diverge. In particular, we will cover the up- and down-sides of direct, offload and symmetric acceleration. GPU and Many Integrated Core devices will be used as examples.
