2

My aim to to explore GPU acceleration for tabular data with 10,000 to 10M+ records. I am most familiar with Pandas, so cuDF seems like a good place to start.

I'm finding mixed results re: whether cuDF will run on my system (Windows 7 Pro 64-bit, i7-6820HQ, 32GB RAM, NVidia Quadro M2000M 4GB). There is also an onboard graphics card.

per the gitHub page (https://github.com/rapidsai/cudf):

CUDA/GPU Requirements

  • CUDA 10.0+ (YES - I have v10.1.120)
  • NVIDIA driver 410.48+ (YES - I have 432.06)
  • Pascal architecture or better (NO - Maxwell)

I have heard that Pascal architecture is preferred/optimal as opposed to a requirement, but maybe that was for older versions of cuDF? Just this morning I heard it will run on Win 64, though performance benefits may also be reduced. Nonetheless, I'm interested in giving it a shot.

When I install from the conda prompt (python 3.6 env) using the recommended command for my CUDA version:

conda install -c rapidsai -c nvidia -c numba -c conda-forge cudf=0.13 python=3.6 cudatoolkit=10.1

I get:

Collecting package metadata (repodata.json): done Solving environment: failed with initial frozen solve. Retrying with flexible solve.

PackagesNotFoundError: The following packages are not available from current channels:

  • cudf=0.13

Current channels:

To search for alternate channels that may provide the conda package you're looking for, navigate to

https://anaconda.org

and use the search bar at the top of the page.

When I go to anaconda.org and search for cuDF (or RAPIDS), all I find are Linux installs.

I attended an Anaconda-sponsored webinar earlier today where the speaker said it'll run in Win-64, though this older post suggest maybe I need to build from source:

Package not found error while installing CuSpatial or CuDf library

I'm not ready to attempt a build from source. Am I just wasting my time? Recommendations appreciated (for either resolving cuDF with my system or alternative packages).

Gonçalo Peres
  • 11,752
  • 3
  • 54
  • 83
CreekGeek
  • 1,809
  • 2
  • 14
  • 24

2 Answers2

7

cuDF maintainer here.

Currently, cuDF nor any other RAPIDS libraries are supported in a native Windows environment. There's an issue tracking Windows support here: https://github.com/rapidsai/cudf/issues/28.

In general, native Windows support is not a priority for us, especially given the push towards GPU support in WSL2 that is currently in open beta.

Keith Kraus
  • 231
  • 1
  • 4
  • 1
    Thank-you for the reply. Lack of Windows support is disappointing, but completely understandable. I really like the approach you guys are taking and have bookmarked it for when I finally getting around to getting a Linux system. Cheers – CreekGeek Aug 27 '20 at 19:44
  • 2
    I disagree that lack of native windows support is completely understandable. Tensorflow and many other GPU-enabled libraries support native Windows, and given that high-end GPUs are still mainly used for gaming, it seems support for the preeminent gaming platform would be more understandable. – csteel Feb 28 '22 at 01:27
4

Apparently there are some news regarding this. Here one can find the guide for using NVIDIA CUDA on Windows Subsystem for Linux.

Getting started with running CUDA on WSL requires you to complete these steps in order:

1. Installing the latest builds from the Microsoft Windows Insider Program

2. Installing the NVIDIA preview driver for WSL 2

3. Installing WSL 2

Important note regarding the installation of the latest builds from the Microsoft Windows Insider Program

Ensure that you install Build version 20145 or higher.

You can check your build version number by running winver via the Windows Run command. (Source)

Hopefully next year a version of Windows that meets the Build version 20145 or higher requirement will be released and then one doesn't need to run an "Insider Program" build.

Source for Windows 10 release information.

Here one will be able to follow all the updates regarding the Support for Windows.

Gonçalo Peres
  • 11,752
  • 3
  • 54
  • 83