rCUDA was presented last September in the HPC Advisory Council Conference that took place in Valencia, Spain. You can watch the video of the talk in in this link
https://www.youtube.com/watch?v=aKNwxE9NNlQ

rCUDA is designed to provide the best performance. In the plot you can see a comparison among the performance of P2P memory copies (data copies among GPUs) carried out with CUDA using the PCIe link within a single node and the performance attained by rCUDA when both GPUs are located in different remote servers and connected by InfiniBand. Three different GPU generations are considered. As you can see, using rCUDA does not mean a performance degradation.

The rCUDA Team is happy to announce that the new rCUDA version, to be released at the end of October 2018, has been successfully tried with the Inception program using TensorFlow 1.10 and CUDA 9.0. More tests need to be carried out with other programs.

 Federico Silla presented rCUDA last week in the High Performance Container Workshop held in conjunction with ISC 2018 in Frankfurt, Germany. You can find a copy of the slides here.

A new release of the rCUDA middleware will be published during September-October. In addition to support CUDA 9.0 and CUDA 9.1, the new version of rCUDA has successfully run on popular HPC applications such as BARRACUDA, CUDAmeme, GPUBlast, GPU-LIBSVM, Gromacs, LAMMPS, MAGMA and NAMD. Deep learning frameworks are also supported. This new version of rCUDA has been successfully run with TensorFlow version 1.7, Caffe, Torch, Theano, PyTorch and MXNET. Finally, renderers such as Blender and Octane are also supported. This new version of rCUDA is a major step forward in the evolution of this middleware because almost 80% of its code is either new or has been reworked.