Running these models requires significant “brain power” from computers, specifically GPUs (Graphics Processing Units). While you can use a basic laptop for simple tasks, fine-tuning a model requires specialized hardware, often accessed through the “cloud” (servers owned by companies like Google, AWS, or Microsoft).
- Pre-training: Requires massive data center-scale GPU clusters (thousands of high-end GPUs). This is done by large organizations like Meta or Google.
- Fine-Tuning: More accessible. You can fine-tune a small model (7B parameters) on a specific task using a single powerful GPU (e.g., NVIDIA A100, V100, or consumer-grade RTX 4090) for a few hours. Techniques like LoRA (which we cover in Module 4) make this even more efficient.
- Inference (Running the model): This can be done on a good CPU for small models, but for speed and responsiveness, a GPU is preferred. Cloud services (like Google Colab, AWS, Azure) offer GPU rentals by the hour, which is a cost-effective way to start.
As an AI practitioner, part of your job is navigating these requirements. You’ll learn to use cloud notebooks, connect to GPU resources, and manage model files—all essential skills for real-world AI projects.