Part I: Models, workload state, and GPU hardware

These chapters establish the work and its storage. Detailed diagnosis follows the execution mechanisms and the first controlled measurement in Part II.

Illustrative token IDs select embedding rows, repeated transformer blocks yield logits, and an int32 tensor example separates shape and byte strides from storage. Training updates weights and saves activations; generation repeats decode with fixed weights and growing KV state. A GPU cutaway separates SM resources, device-wide L2, HBM, host transfers, peers and networking.
Figure 1: Model operations create tensor work; training and generation keep different live state, and GPU storage and links serve different scopes.

Chapters in this part