AMD released its Radeon Open Compute platform (ROCm) 1.7 release as part of their wares at this week's Supercomputing 17 (SC17) conference in Denver.
The ROCm 1.7 update introduces multi-GPU support for "the latest Radeon GPU hardware" (presumably referring to Vega) while also supporting TensorFlow and Caffe via AMD's MIOpen libraries as Phoronix reported today:
It also looks like there may be some more performance optimizations in ROCm 1.7, but their press email is light on technical details. As of writing, the ROCm documentation hasn't been updated for ROCm 1.7 but that should shed some more light on this new release when it surfaces, probably later in the day. If the changes look interesting enough, I'll be working on some fresh ROCm OpenCL benchmarks. For now at least through Linux 4.15, ROCm continues to depend on kernel changes not yet mainlined and in user-space against LLVM/Clang/LLD.
AMD is also using the SC17 conference for announcing the broader availability of AMD EPYC systems by a variety of companies.