Skip to main content

1 - Create a GPU Instance on Google Cloud

Why this guide's resources? This guide builds the cloud side of the tutorial: the CloudBank billing account gets you into your Google Cloud project, Cloud Shell gives you an authenticated gcloud terminal with nothing to install, and the Compute Engine GPU VM (1x T4, n1-standard-4, Ubuntu 22.04 LTS, 200 GB pd-ssd) is the machine that will run the simulation. See 0 - The Cloud Resources in This Tutorial for the full resource map.

Goal: Create a GPU-enabled virtual machine instance on Google Cloud Platform to run GPU-accelerated Julia ocean modeling simulations.

Prerequisites​

  • CloudBank will create the Google account for you automatically.
  • A desktop browser. On Chromebooks, open an Incognito window first to ensure you use the CloudBank-generated Google account.

Login to the Google Cloud Console​

  1. Open the login page: https://www.cloudbank.org/billing-account-access (Incognito if on Chromebook).
  2. In the table under Public Cloud Web Console Login, click the Login button in the Google Cloud Platform row.
  3. When prompted, continue with the Google account generated by CloudBank (no need to supply your own).
  4. You should land in the Google Cloud Console for your assigned project. In the top bar, note the active project name/ID; we will need it for creating the VM instance.

Enable Compute Engine API​

  1. Open the Compute Engine page: Navigation menu ▸ Compute Engine ▸ VM instances.
  2. If prompted, click ENABLE to turn on the Compute Engine API. Wait until the page refreshes.

Connect with Cloud Shell​

We'll create the GPU instance using Gemini AI assistant in Cloud Shell. Open Cloud Shell by clicking the terminal icon at the very top of the screen labeled Activate Cloud Shell.

Create the GPU instance with Gemini​

Step 1: Enable Compute Engine API​

  1. In Cloud Shell, type gemini to activate the Gemini AI assistant
  2. Ask Gemini to enable the Compute Engine API:
I need to enable the Compute Engine API. Please run the command to enable it for me.

Step 2: Verify GPU availability​

Ask Gemini to check GPU availability:

I need to create a GPU instance with NVIDIA Tesla T4. Please check if the T4 GPU is available in zone us-west1-b by listing the accelerator types.

Step 3: Create the GPU instance​

Ask Gemini to create the instance with these specifications:

I need to create a GPU instance with:
- Name: julia-ocean-gpu
- Zone: us-west1-b
- Machine type: n1-standard-4
- GPU: nvidia-tesla-t4 (1 GPU)
- Image: ubuntu-2204-lts (specify with --image-family)
- Boot disk: 200GB SSD
- Maintenance policy: TERMINATE

Please create this instance using gcloud compute instances create.

Image flag: use --image-family=ubuntu-2204-lts in the creation command (not --image=ubuntu-2204-lts). On current gcloud (verified with gcloud 587.0.0, October 2026), --image=ubuntu-2204-lts no longer resolves image family names and fails with "referenced image resource cannot be found"; --image-family is the flag that resolves the family, and it is what Gemini uses above.

Do NOT ask Gemini to enable the install-nvidia-driver=True metadata flag. On Ubuntu 22.04 that flag sets the metadata but does not install a driver (verified: after boot there was no nvidia-smi and no NVIDIA packages). We install the driver explicitly in Step 5 below.

Expected Gemini response:

The instance julia-ocean-gpu has been created successfully in us-west1-b.

I've configured the instance with the following specifications:
 - Machine Type: n1-standard-4 (x86)
 - GPU: 1x nvidia-tesla-t4
 - Image: ubuntu-2204-lts (x86_64, specified with --image-family; resolved to ubuntu-2204-jammy-v20261016). Note that I used the x86 version instead of arm64 because the T4 accelerator and N1 machine family are x86-based.
 - Boot Disk: 200GB SSD (pd-ssd).
 - No NVIDIA driver was installed; we will install it with explicit commands in Step 5.

Note: This creates:

  • An Ubuntu 22.04 LTS x86_64 instance with 4 vCPUs and 15 GB RAM
  • One NVIDIA Tesla T4 GPU (16 GB VRAM)
  • 200 GB SSD storage for data and packages
  • No NVIDIA driver yet (installed in Step 5 below)

Step 4: Verify instance status​

Ask Gemini to check the instance status:

Please verify that the instance julia-ocean-gpu in zone us-west1-b is running and show me its details including GPU status.

You should see the instance status as RUNNING with the T4 accelerator attached.

Note: If you ask Gemini to run gcloud compute accelerator-types list to check T4 availability, it may crash in Cloud Shell with a Python error (AttributeError: 'NoneType' object has no attribute 'GetRequestType'). This is a known gcloud CLI bug in the accelerator-types command; T4 availability is confirmed implicitly by the successful creation of the instance with the GPU attached, so you can skip that check.

Step 5: Install NVIDIA drivers​

Ask Gemini to install the NVIDIA driver over SSH:

SSH into the server and install the NVIDIA driver. First run apt-get update, then install the nvidia-driver-535 package, then reboot the server. After the reboot, reconnect and check nvidia-smi.

Installation commands (run inside the instance):

sudo apt-get update
sudo apt-get install -y nvidia-driver-535
sudo reboot

Wait for the reboot to finish, then reconnect:

gcloud compute ssh julia-ocean-gpu --zone=us-west1-b

Note: On Ubuntu 22.04 the --metadata=install-nvidia-driver=True option is NOT sufficient by itself: it only records the metadata flag and does not install a driver (verified live: no NVIDIA packages, no nvidia-smi after boot). The explicit apt install + reboot above is the reliable path, and it is what this tutorial uses. nvidia-driver-535 is the current driver package name on Ubuntu 22.04 LTS but the exact version apt delivers may be newer than 535.x (verified October 2026: it installs driver 580.178.04 / CUDA 13.0). The exact driver version does not matter for the tutorial: CUDA.jl ships its own runtime artifacts, and any 500-series-or-newer driver works.

Step 6: Verify GPU installation​

Ask Gemini to check the GPU status on the server:

Show me the output of nvidia-smi on the server.

Expected Gemini response:

I'll check the nvidia-smi output for you.

[... shows nvidia-smi output with Tesla T4 GPU, 15360 MiB memory, Driver Version: 580.178.04, CUDA Version: 13.0 ...]

This confirms:

  • GPU: Tesla T4 detected
  • Driver installed and running
  • 15360 MiB (~15 GB) GPU memory available
  • CUDA Version (max supported by driver): 13.0. This is new enough for CUDA.jl: its toolchain reports runtime 13.2 / driver 580.178.4 inside Julia (see the expected CUDA.versioninfo() output in tutorial 2).

Step 7: Install Julia​

Ask Gemini to install Julia using Juliaup:

SSH into the server and install Julia using Juliaup. First run the curl command to install Juliaup, then add Julia version 1.11 and set it as default.

Julia installation commands:

curl -fsSL https://install.julialang.org | sh
source ~/.bashrc
juliaup add 1.11
juliaup default 1.11

Step 8: Verify Julia installation​

Ask Gemini to check the Julia version:

Show me the Julia version on the server.

Expected output: julia version 1.11.x

Step 9: Install Jupyter​

Ask Gemini to install the Jupyter notebook server:

SSH into the server and install the Jupyter notebook server via apt.

Installation commands:

sudo apt-get install -y jupyter-nbconvert jupyter-notebook

The jupyter-notebook package provides the notebook server (version 6.4.8 on Ubuntu 22.04); jupyter-nbconvert alone is NOT enough because IJulia's notebook() command launches the notebook server and needs its notebook subcommand. We register the Julia kernel in the next step, after IJulia is installed (the kernel registration needs the IJulia package, so it must come after the Julia package install below).

Step 10: Prepare the Simulation Environment​

Ask Gemini to clone the repository and install all Julia packages:

SSH into the server, clone the showcase repository, then install all required Julia packages globally with the correct versions. After the package install, register the Julia kernel for Jupyter.

Commands:

git clone https://github.com/cloudbank-project/cloudbank_showcase.git
cd cloudbank_showcase/gpu_computing_oceanography/

Then in the Julia REPL (or via julia -e '...'):

using Pkg

Pkg.add([
    PackageSpec(name="Oceananigans", version="0.99.0"),
    PackageSpec(name="Reactant", version="0.2.269"),
    PackageSpec(name="Enzyme"),
    PackageSpec(name="CUDA"),
    PackageSpec(name="CairoMakie"),
    PackageSpec(name="ClimaOcean"),
    PackageSpec(name="SeawaterPolynomials"),
    PackageSpec(name="IJulia"),
    PackageSpec(name="JLD2"),
    PackageSpec(name="MPI"),
    PackageSpec(name="FileIO"),
])
Pkg.build("CUDA")

Reactant version pin: Reactant is pinned to 0.2.269, the last release compatible with Oceananigans 0.99.0. The newest Reactant release (0.2.295, October 2026) breaks this tutorial at the GPU time-step with an unadapted traced value kernel-argument error (_check_no_traced_in_kernel_arg, ReactantCUDAExt.jl). Do not install Reactant unpinned.

Register the Julia kernel (after the package install):

julia -e "using IJulia; installkernel(\"Julia\")"

Important: Packages are installed into the global environment; do not run Pkg.activate(".") before this step. Precompilation will take 20–25 minutes on first run.

Expected Gemini response:

The Julia environment on julia-ocean-gpu is now fully set up.

Summary:
 - Project Cloned: ~/cloudbank_showcase/gpu_computing_oceanography
 - Packages installed globally: Oceananigans v0.99.0, Reactant v0.2.269, Enzyme, CUDA, etc.
 - CUDA backend rebuilt and verified.
 - Julia kernel registered for Jupyter.

Important Note: Oceananigans emitted a warning about being tested primarily on Julia v1.10, but it successfully precompiled for your v1.11.9 installation.

Next steps - Choose your access method​

Now your GPU instance is ready! Choose one of these methods to access the Jupyter notebook:

Option 1: Browser-based Jupyter (Tutorial 2)

Option 2: VS Code Remote (Tutorial 3)

  • Full IDE experience with debugging and IntelliSense
  • See tutorial: 3_access_jupyter_vscode.md
  • Best for: Serious development, debugging, advanced users

Both methods provide full access to the GPU-accelerated Julia environment and the ocean modeling notebook.

Cleanup​

Remember to delete the instance when you're done to avoid ongoing charges:

# Run this from Cloud Shell (not inside the instance)
gcloud compute instances stop julia-ocean-gpu --zone=us-west1-b
gcloud compute instances delete julia-ocean-gpu --zone=us-west1-b