Backwards Compatibility Issues (CUDA 9.1/CUDA 8.0)

This is directed to NVIDIA: I have a NVIDIA GeFOrce GTX 1080 Ti and built my system recently to run my own Deep Learning experiments; it runs Windows 10 Professional. However as it turns out that CUDA 8.0 is no longer supported by the hardware (The installer says that the hardware is too new for it, and also that CUDA 8 is not supported by Visual Studio 2017). That’s fine; I proceeded to install CUDA 9.1, which installed perfectly (and also works with Visual Studio 2017, which is a good thing), but as it turns out that none of the frameworks (Tendorflow/Theano/CNTK etc) have support for CUDA 9.1 yet and as such I am unable to work with my DL codes (which use tensorflow libraries on top of Keras).

My question is why don’t you guys introduce backwards compatibility in these things? Why do you not make it so that the latest versions of CUDA/cuDNN also support everything related to the older versions (for example CUDA 9.1 supporting everything based off of CUDA 9 and earlier). This would solve a lot of issues that users face over time in regards to upgrading/downgrading (which is a pain) every-time a new version is released/ going back and forth from a newer version to an older one etc etc.

Yup, I second that.

But, looking forward you might find at least Tensorflow in newer builds supporting cuda 9.1 – and possibly delivering a 30% speed bump (your mileage will vary ;)