Compute is a large selling point of GCN. Typically the design has favored OpenCL and other compute work, it’s been very efficient at it. That is to say that AMD’s current GPU designs are very good at parallelization and is able to use the resources at hand effectively. We have a few different benchmarks to demonstrate this, the first being an actual DNN. Caffe DNN can easily be run by anyone on a *nix distribution. Simply download the applicable binaries, compile it with either cuDNN or use AMD’s open source OpenCL version. You can then test the benchmark that’s inherent in the Caffe framework.
Compute remains a very important purview of AMD’s GPUs. Though Polaris doesn’t quite have the DP chops that Hawaii, the reining double-precision champion, it’s still decent enough to be a very good and inexpensive solution for scientific compute projects. It slots right in with our projections and is fast enough to do a lot, and especially when combined in parallel with more GPUs, the RX 480 is pretty capable.