GPUs have become an increasingly popular general computing platform. However, their efficient utilization requires some understanding of the architecture and programming model. A sensible question arises about their applicability in terms of the necessary time investment versus
the added scientific value.

An introduction on the basics of GPGPUs is presented first, including the architecture, the computing model and a general methodology for parallelizing an algorithm.
Opportunities and pitfalls are then elucidated with basic examples of problems/implementations suited or unsuited for GPUs. Assessing a problem
for its applicability to GPUs is shown as a key decision-making step.

The seminar is not only envisioned as an introduction to GPU programming; key concepts are elucidated specifically on a concrete case of Finite Element (FE)
algorithm parallelization. Concepts such as problem-breakdown, testing, verification, profiling and optimization are presented through the FE algorithm's implementation.

Overall, a realistic assessment of adopting GPU computation (and everything it can entail) is provided through the lens of added scientific value.

Detailed time/place info

08.05.2017, 10:00-12:00