How to overcome the huge increase of windows dlls - minimum cuda/cudnn builds that support Ampere?

I’m upgrading from Cuda 10.x to 11.x in order to support Ampere cards for a Windows application.
However not only the required cublas/cublasLt dlls went through a x5 size increase and still grow by tens of megabytes with each update but the cuddn*infer.dll also increased drastically in size.
I suppose that the trend is to optimize by hardcoding thousands of specialized kernels but this has really gone out of control!


I plotted the total size of biggest files I need to redistribute for an OpenCV based application.
It went from 152MB (7zipped) to 389MB!

So I was wondering

  1. what is the mimimum package that supports compute_86? I was expecting support from Cuda 11.0 but nvcc rejects anything above compute_80. Is Cuda 11.1 the minimum
  2. are there any problems to expect when using older builds? say Cuda 11.2.2 w/ cudnn 8.1.1

Any change to get an official reply on support of Ampere in Cuda 11.0/11.1? @jhalabi maybe?

Hi,

From CUDA 11.0 onwards support Ampere.
Please refer to the support matrix doc to get more info.

Could you please let us know which steps are you following to upgrade CUDA. Have you tried uninstalling the old version and installing the new version again. Did it solve the problem?

@spolisetty thank you for your reply
I have tried 11.0, 11.1 up to 11.7
With the latter: nvcc rejects compiling with sm_86, looking at the help

Have I done anything wrong @spolisetty ?

Hi,

Please make sure you’re following the correct installation steps by following the official doc and also please make sure CUDA is installed correctly.

Please uninstall the old package completely and try a fresh installation.

Thank you.

@spolisetty I will try once again.
Looking carefully at the release notes, cuda 11.0 adds partial Ampere compatibility (sm_80) Release Notes :: CUDA Toolkit Documentation

CUDA 11.0 adds support for the NVIDIA Ampere GPU microarchitecture (compute_80 and sm_80).

sm_86 requires at least cuda 11.1: Release Notes :: CUDA Toolkit Documentation

Added support for NVIDIA Ampere GPU architecture based GA10x GPUs GPUs (compute capability 8.6), including the GeForce RTX-30 series.

@spolisetty Since my last message, I’ve seen in the release notes that Nvidia wanted to better split dlls in order to mitigate the size increase. I also so that release 8.5.0 had slightly smaller dlls.
However today I gave a look at cuda 11.8 + cudnn8.7.0 and again it’s a huge increase over previous version.
Is there any official word addressing that issue? Because redistributing programs with cuda supports becomes more and more difficult due to the size of the libraries