Hello Model
← Training basics

Lesson 6 of 8

CPU vs GPU, and keeping costs down

In plain words: Tabular models train fine on a laptop CPU; images, text, audio and LLMs usually need a GPU.
  • CPU is enough for: tabular data, classical ML, small text models with few examples.
  • GPU helps for: fine-tuning transformers, vision and speech models, self-hosting LLMs.
  • Start free: Google Colab and Kaggle notebooks include GPUs.
  • Rent GPUs by the hour for training, then serve on CPU with a smaller or quantised model where possible.
  • Use spot/preemptible instances for training jobs — often 60–90% cheaper.
Tip Turn off idle notebooks and endpoints. Forgotten GPUs are the most common surprise bill.