Looks like it's on pypi now https://pypi.python.org/pypi/tensorflow-gpu

I thought the plan was to have common code for the generic, non-critical parts like graph handling, while operators would be shipped in multiple variants, keyed and loaded at runtime by CPUID. You can even delegate that to glibc if they are packaged as shared libraries and it's a recent enough version. Maybe in 1.7?

Do you have any evidence that the TF developers are going to do that?

I heard it from someone at Google at a public event, but there was no date attached to it. I should have asked.