Skip to content

Cloud GPU setup

Don't have a GPU or sufficient resources on your local machine? For approximately €0.17-1.80 per hour (depending on GPU model), you can rent GPU instances from cloud providers to run GOLIAT simulations. This guide walks you through setting up a cloud Windows VM with GPU support.

Overview

This setup allows you to:

  1. Deploy a Windows VM with GPU support (from RTX 4090 to H100) via cloud providers
  2. RDP into the machine remotely
  3. Automatically install and configure everything needed to run GOLIAT
  4. Launch simulations with minimal manual intervention

The entire setup process takes approximately 10 minutes and is fully automated once you copy the setup script to the VM.

Prerequisites

  • A cloud provider account (this guide uses TensorDock as an example, but the process is similar for other providers)
  • Basic familiarity with Remote Desktop Protocol (RDP)
  • Access to Sim4Life installer and license (these are downloaded automatically or can be provided manually)

Step 1: Deploy a GPU instance

Option A: Using the web interface

  1. Visit TensorDock Dashboard
  2. Select your GPU model (e.g., RTX 4090)
  3. Configure resources:
  4. CPU: 8+ cores recommended
  5. RAM: 32 GB+ recommended
  6. Storage: 250 GB+ recommended
  7. OS: Windows 10 or Windows 11
  8. Prefer Dedicated IP if available
  9. Set a secure password for the VM
  10. Deploy the instance

Option B: Using the Python script

A Python script is provided in cloud_setup/deploy_windows_vm.py to automate VM deployment via the TensorDock API:

:: From the repository root:
copy .env.example .env
:: Set TENSORDOCK_API_TOKEN and TENSORDOCK_VM_PASSWORD in .env.

python cloud_setup\deploy_windows_vm.py

The same .env file is used by gpu_dashboard.py. It is ignored by Git; .env.example documents the supported variables without containing credentials. Rotate a token immediately if it is ever committed or printed in a log.

Step 2: Connect via RDP

Once your VM is deployed:

  1. Find the public IP address in your provider's dashboard
  2. Use Windows Remote Desktop Connection (or any RDP client)
  3. Connect using:
  4. IP: The public IP from the dashboard
  5. Username: Usually Administrator or user (check provider docs)
  6. Password: The password you set during deployment

Step 3: Run the setup script

The cloud_setup/ directory contains a unified setup script that handles both fresh installation and reconnection scenarios.

What the script does

The setup.bat script automatically detects whether GOLIAT has already been installed by checking if the goliat/ folder exists in the user's home directory.

Fresh installation (goliat/ not found)

If this is a new machine, the script runs the full setup flow:

  1. Checks computer name and administrator privileges
  2. Downloads and installs OpenVPN
  3. Downloads and installs Python 3.11
  4. Installs gdown utility for Google Drive downloads
  5. Downloads and installs Sim4Life
  6. Downloads VPN configuration files
  7. Connects to VPN (if required for Sim4Life license access)
  8. Installs Git and clones the GOLIAT repository
  9. Automatically configures Sim4Life license (enters license server, validates, and completes installation)
  10. Git Bash automatically runs initialization commands (pip install, git config, goliat init)

Reconnection (goliat/ found)

If the setup has already been completed, the script runs the reconnection flow:

  1. Connects to VPN
  2. Opens File Explorer at the goliat/ directory
  3. Launches Git Bash with automatic commands:
  4. Sets git safe.directory
  5. Configures git user.email and user.name
  6. Runs git pull to fetch latest changes
  7. Leaves the terminal ready for use

This means you only need one script - just run setup.bat every time you connect to the VM.

Running the script

  1. Copy cloud_setup/setup.bat to the VM.

  2. Copy .env.example to .env, fill the GOLIAT_* bootstrap settings and SIM4LIFE_LICENSE_SERVER, and place that .env beside setup.bat. Do not edit credentials into the batch file. The local dashboard sends a filtered bootstrap .env automatically and never sends the TensorDock API token.

  3. Run as Administrator:

Right-click setup.bat → Run as administrator

  1. First run: Wait ~10 minutes while the script installs everything. License configuration is automated (takes 30-60 seconds for validation). If VPN is not connected, the license step will fail with a clear error message - you can configure it manually later.

  2. Subsequent runs: The script detects the existing installation and runs the quick reconnection flow (~15 seconds).

File structure

The cloud_setup/ directory contains:

  • setup.bat: Unified setup script for fresh installation and reconnection; reads settings from a neighboring .env
  • license_automation.py: Automates Sim4Life license configuration, accepts --license-server argument
  • deploy_windows_vm.py: Python script for API-based VM deployment; reads credentials from .env
  • gpu_dashboard.py: Local VM dashboard; reads TENSORDOCK_API_TOKEN from .env

Do not create parallel credential-bearing source copies. Keep all local values in .env, which is ignored by Git. Delete the VM copy when it is no longer needed, and restrict access to the VM account because it contains bootstrap credentials.

Cost estimation

TensorDock pricing varies by GPU model and location, ranging from approximately €0.17-1.80/hour:

Enterprise GPUs: H100 (~€1.80/hr), A100 (~€0.78/hr), L40S (~€0.60/hr), Tesla V100 (~€0.19-0.27/hr)

Workstation RTX GPUs: RTX PRO 6000 (~€0.91/hr), RTX 6000 ADA (~€0.64/hr), RTX 5090 (~€0.45/hr), RTX A6000 (~€0.36/hr), RTX 4090 (~€0.32/hr), RTX 3090 (~€0.18/hr)

Notable features:

  • High-memory configurations available (up to 512 GB RAM)
  • Per-second billing with no minimum commitment
  • Instances can be stopped when not in use to minimize costs

Monthly estimate (assuming 730 hours): ~€15-130/month for running instances, plus storage costs when stopped.

For current pricing and availability, consult TensorDock's dashboard directly.

Provider comparison

Hyperscalers (GCP, Azure, AWS): Can create VM images/snapshots and duplicate them to new instances, but quota requests for GPU instances can take days or weeks to be approved. Generally higher pricing.

Specialized GPU providers (e.g., TensorDock): No quota requests - instant GPU access. Lower pricing (typically 20-40% cheaper than hyperscalers). Image duplication feature may be available depending on provider.

Alternative providers

While this guide focuses on TensorDock, similar setups work with AWS EC2 (G4/G5 instances with Windows Server), Google Cloud Platform (GPU-enabled Windows VMs), and Azure (NV-series VMs with Windows). The setup script may need minor modifications for different providers (e.g., different default usernames, network configurations).

Headless / SSH access

For managing the VM programmatically from a Linux host, see cloud_setup/ssh/README.md. It documents a one-shot script to enable key-only OpenSSH on the VM (reusing the Tensordock :8888 port forward, locked-down ACL on administrators_authorized_keys, elevated admin tokens for SSH sessions), plus the conventions for running goliat study from SSH.

  • oSPARC: Cloud batch execution via oSPARC platform (alternative to VM setup)
  • Monitoring dashboard: When running studies across multiple cloud VMs, use the monitoring dashboard to track progress, view logs, and coordinate super studies across all workers
  • SSH bring-up: Manage the VM from a Linux host