AWS Compute Selection Guide for Distributed Training (Part 1: Model Scale and Hardware Selection)

When preparing to train a large language model (LLM), the most critical decision goes beyond simply choosing a GPU model — it's about building infrastructure that matches the model's scale and training approach.

You need to decide whether a single node is sufficient or whether you need to scale to multiple nodes, and the method of resource acquisition is also a consideration. Specifically, you should review the following options.

• On-Demand: Use immediately when needed
• Capacity Block: Reserve resources in advance for a specific period
• Amazon EC2 UltraClusters: High-performance cluster configuration for large-scale distributed training

For efficient training, you need an optimal AWS compute strategy that considers not just hardware specs but also network performance and inter-node scalability.