A virtual machine (VM) is a software-defined computer that runs an operating system and applications in an isolated environment. For large-model training, it provides access to powerful multi-GPU hardware without purchasing the physical rigs. section 5.1.1
Using a cloud VM involves practical trade-offs:
- Renting multi-GPU capacity is usually cheaper than buying equivalent hardware, but hourly charges continue until the environment is stopped. section 5.1.1
- Google Compute Engine offers a straightforward route to a multi-GPU environment after account, project, billing, and gcloud setup. section 5.1.1
- Smaller models can run locally or on single-GPU free tiers, while larger models may justify faster, more efficient multi-GPU training. section 5.1.1
Create and remove a training VM
A typical workflow creates a Google Cloud instance with a GPU image, connects over SSH, copies application files, and deletes the instance when finished. The final step prevents unnecessary charges.
