GPU support in the current Auryn versions had been discontinued, because it did not improve performance in most use cases, but rather degraded it. See http://journal.frontiersin.org/article/10.3389/fninf.2014.00076/full for details.
In particular I tried to integrate neuronal state variables on GPU while spike propagation and plastic updates still happened on the CPU. The constant back and forth between the two killed all performance gain from the GPU. The root of all evil why plasticity and spike propagation cannot easily be implemented efficiently on the GPU are the sparse connectivity matrices used in most network models. If connectivity were dense, an efficient implementations would be much easier, because both neuronal state updates *and* spike propagation (plus backward spike propagation for STDP) could be implemented more efficiently on GPUs. Other simulators such as Brian have been fighting with similar problems.
If you would like to try your luck, a version of IFGroup written in OpenCL is available upon request from the author.
— Friedemann Zenke 2016/06/01 18:35