WSL 2 GPU Support for Docker Desktop on NVIDIA GPUs

It’s been a year since Ben wrote about Nvidia support on Docker Desktop. At that time, it was necessary to take part in the Windows Insider program, use Beta CUDA drivers, and use a Docker Desktop tech preview build. Today, everything has  changed:

  • On the OS side, Windows 11 users can now enable their GPU without participating in  the Windows Insider program. Windows 10 users still need to register.
  • Nvidia CUDA drivers have been released.
  • Last, the GPU support has been merged in Docker Desktop (in fact since version 3.1).

Nvidia used the term near-native to describe the performance to be expected.

Where to find the Docker images

Base Docker images are hosted at The original project is located at

What they contain

The nvidia-smi utility allows users to query information on the accessible devices.

$ docker run -it --gpus=all --rm nvidia/cuda:11.4.2-base-ubuntu20.04 nvidia-smi
Tue Dec  7 13:25:19 2021
| NVIDIA-SMI 510.00       Driver Version: 510.06       CUDA Version: 11.6     |
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|   0  NVIDIA GeForce ...  On   | 00000000:01:00.0 Off |                  N/A |
| N/A    0C    P0    13W /  N/A |    132MiB /  4096MiB |     N/A      Default |
|                               |                      |                  N/A |

| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|  No running processes found                                                 |

The dmon function of nvidia-smi allows monitoring the GPU parameters :

$ docker exec -ti $(docker ps -ql) bash
root@7d3f4cbdeabb:/src# nvidia-smi dmon
# gpu   pwr gtemp mtemp    sm   mem   enc   dec  mclk  pclk
# Idx     W     C     C     %     %     %     %   MHz   MHz
    0    29    69     -     -     -     0     0  4996  1845
    0    30    69     -     -     -     0     0  4995  1844

The nbody utility is a CUDA sample that provides a benchmarking mode.

$ docker run -it --gpus=all --rm nbody -benchmark
> 1 Devices used for simulation
GPU Device 0: "Turing" with compute capability 7.5

> Compute 7.5 CUDA device: [NVIDIA GeForce GTX 1650 Ti]
16384 bodies, total time for 10 iterations: 25.958 ms
= 103.410 billion interactions per second
= 2068.205 single-precision GFLOP/s at 20 flops per interaction

Quick comparison to a CPU suggest a different order of magnitude of performance. GPU is 2000 times faster:

> Simulation with CPU
4096 bodies, total time for 10 iterations: 3221.642 ms
= 0.052 billion interactions per second
= 1.042 single-precision GFLOP/s at 20 flops per interaction

What can you do with a paravirtualized GPU?

Run cryptographic tools

Using a GPU is of course useful when operations can be heavily parallelized. That’s the case for hash analysis. dizcza hosted its nvidia-docker based images of hashcat on Docker hub. This image magically works on Docker Desktop!

$ docker run -it --gpus=all --rm dizcza/docker-hashcat //bin/bash
root@a6752716788d:~# hashcat -I
hashcat (v6.2.3) starting in backend information mode


CUDA Info:

CUDA.Version.: 11.6

Backend Device ID #1
  Name...........: NVIDIA GeForce GTX 1650 Ti
  Processor(s)...: 16
  Clock..........: 1485
  Memory.Total...: 4095 MB
  Memory.Free....: 3325 MB
  PCI.Addr.BDFe..: 0000:01:00.0

From there it is possible to run hashcat benchmark

hashcat -b
Hashmode: 0 - MD5
Speed.#1.........: 11800.8 MH/s (90.34ms) @ Accel:64 Loops:1024 Thr:1024 Vec:1
Hashmode: 100 - SHA1
Speed.#1.........:  4021.7 MH/s (66.13ms) @ Accel:32 Loops:512 Thr:1024 Vec:1
Hashmode: 1400 - SHA2-256
Speed.#1.........:  1710.1 MH/s (77.89ms) @ Accel:8 Loops:1024 Thr:1024 Vec:1

Draw fractals

The project at uses CUDA for generating fractals. There are two steps to build and run on Linux. Let’s see if we can have it running on Docker Desktop. A simple Dockerfile with nothing fancy helps for that.

# syntax = docker/dockerfile:1.3-labs
FROM nvidia/cuda:11.4.2-base-ubuntu20.04
RUN apt -y update
RUN DEBIAN_FRONTEND=noninteractive apt -yq install git nano libtiff-dev cuda-toolkit-11-4
RUN git clone --depth 1 /src
RUN sed 's/4736/1024/' -i # Make the generated image smaller
RUN make

And then we can build and run:

$ docker build . -t cudafractal
$ docker run --gpus=all -ti --rm -v ${PWD}:/tmp/ cudafractal ./fractal -n 15 -c test.coeff -m -15 -M 15 -l -15 -L 15

Note that the --gpus=all is only available to the run command. It’s not possible to add GPU intensive steps during the build.

Here’s an example image:


Machine learning

Well really, looking at GPU usage without looking at machine learning would be a miss. The tensorflow:latest-gpu image can take advantage of the GPU in Docker Desktop. I will simply point you to Anca’s blog earlier this year. She described a tensorflow example and deployed it in the cloud:

Conclusion: What are the benefits for developers? 

At Docker, we want to provide a turn key solution for developers to execute their workflows seamlessly:


Join us for DockerCon2022 on Tuesday, May 10. DockerCon is a free, one day virtual event that is a unique experience for developers and development teams who are building the next generation of modern applications. If you want to learn about how to go from code to cloud fast and how to solve your development challenges, DockerCon 2022 offers engaging live content to help you build, share and run your applications. Register today at


0 thoughts on "WSL 2 GPU Support for Docker Desktop on NVIDIA GPUs"