You already know: Adam keeps two moment estimates per parameter (D2 Optimization); mixed precision and the 16 bytes/parameter of training state (D5 LLM pretraining, D14 bridge); BF16 vs FP16 range (D12 Quantization); data parallelism and all-gather/reduce-scatter (Parallelism strategies, Collective communication).
Why this matters for Atlas. Training memory has four parts: parameters, gradients, optimizer state and activations. Data parallelism replicates the first three on every GPU, wasting most of the cluster's memory on identical copies. ZeRO and FSDP shard them instead, so N GPUs hold one copy between them. Activation checkpointing trades recomputation for activation memory. Mixed precision halves bytes without losing accuracy. Together they decide the largest model and batch you can train on given hardware, often by 8× or more. This lesson derives each saving, implements sharded training in NumPy to show it's exact, and quantifies the compute and communication each technique costs.
1. Where the memory goes
The question. What exactly fills a GPU during training?
Start from nothing. With mixed-precision training and Adam, per parameter: