Reviving CNTK with CUDA 11.4 for Ampere support
In this blog post I quickly cover how I’ve revived the open source Microsoft Cognitive Toolkit (CNTK) with CUDA 11.4 (and cuDNN 8.2) in order to get support for NVidia Ampere/30 series GPUs.
This work is based on the community effort discussed in
Is anyone trying to build CNTK with CUDA 11.1?
and not the least the branch
by @haryngod who did the hard
work of fixing code due to breaking API changes in CUDA libraries.
I, however, did the work in incremental steps incl. reverse engineering nuget packages so new ones could be made. These steps were done as a series of pull requests (PRs) in my fork of the repository as detailed below.
- Building with Visual Studio 2019 and CUDA 10.2 (incl. patches) - this PR updates the library to use CUDA 10.2 (from 10.0) to be able to build CNTK with as few changes as possible with Visual Studio 2019. However, still with the C++ v141 toolset. The main idea here was to verify that I could build CNTK and that it worked as expected in pipelines at my work.
- Add SWIG generated wrapper C++ and C# code - this PR adds the SWIG generated code for C# bindings to git. This is perhaps controversial but having this in git makes it much easier to inspect, debug etc.
- Add NuGet package definitions (reverse engineered) -
this PR adds
nuspecfiles and similar reverse engineered from existing nuget packages for CNTK 2.7.0. This marks the first release of my fork with the first nuget packages as Release 2.7.1.
- CUDA 11.1.1 - building upon
2.7-cuda-11.1branch this PR switches to using CUDA 11.1.1. There were several other changes needed to get this to compile. This was released as Release 2.8.0.
- CUDA 11.4 - finally I update to the latest available CUDA version (11.4) at this time. This required a lot of changes, due to CUDA apparently doing more template instantiations with increasing version, requiring adding overloads of several low level methods before the code would compile. This was released as Release 2.8.1.
All the work was done on and for Windows x86-64 using Visual Studio 2019 (
and was done as quickly as possible. Only single GPU is supported.
I only verified this works with the C# ML pipelines we have at my work, so there are no guarantees this will work for your workloads.
Note that due to code generation changes that output will change depending on which GPU you have. Our pipelines are fully reproducible (to the bit) on the same GPU, but we observed minor changes in output on 20 series GPUs and of course on Ampere GPUs. For the 10 series GPUs there were no changes. Reproducibility in CNTK is ensured with:
1 2 CNTKLib.ResetRandomSeed(fixedRandomSeed); CNTKLib.ForceDeterministicAlgorithms()
Floating point output changes are expected if binary code changes. Same applies if the C++ compiler changes code generated for floating points on CPUs. Order of instructions etc.
Each of the above releases contains the nuget packages too. I very much wanted to
publish these on nuget.org, but they have a 250MB limit on
package sizes, which is impossible to hit since e.g. the file
cudnn_cnn_infer64_8.dll in cuDNN has a size of
~615MB, which cannot be
compressed below 250MB with zip. This limit is mentioned in
Yes, the limit is 250 MB. However, it’s extremely rare for packages to be that large. In fact, less than 3% of packages are larger than 10 MB (less than 0.1% are larger than 100 MB).
Just because something is rare, does not mean there isn’t a need for larger packages. And if they are rare the impact on overall disk usage should be small, so it is a bit counterintuitive.
GitHub Package Registry like Azure Artifacts appears to support packages up to 512MB, so I have published the packages on my GitHub packages feed at https://github.com/nietras?tab=packages. However, GitHub does not allow anonymous un-authenticated access to this package feed, which means I would have to distribute an API key for read accesss to this, but users would not easily be able to add this to their git repositories (e.g. on GitHub) since such tokens are automatically blocked. This is covered in Download from Github Package Registry without authentication and has not yet been resolved.
That’s all. If you end up using any of the above releases and encounter any issues feel free to create a pull request to my fork and I can do a new release at some point.