EECS Publication
On block-asynchronous execution on GPUs
Hartwig Anzt and Jack Dongarra and Edmond Chow
This paper experimentally investigates how GPUs execute instructions when used for general purpose computing (GPGPU). We use a light-weight realizing a vector operation to analyze which vector entries are updated subsequently, and identify regions where parallel execution can be expected. The results help us to understand how GPUs operate, and map this operation mode to the mathematical concept of asynchronism. In particular it helps to understand the effects that can occur when implementing a fixed-point method using in-place updates on GPU hardware. Keywords: GPU-computing, asynchronous execution, block-asynchronous iteration
Published 2016-11-18 05:00:00 as ut-eecs-16-746 (ID:606)