Preprint / scholarly article
I/O-First Energy Reduction for Transformer-Scale AI
- Published
- DOI
- 10.5281/zenodo.18015237
Abstract
The preprint treats boundary I/O bytes and peak hot memory as auditable budgets, defines an I/O Gain-Shut Kernel that enforces fail-closed admission for transfers and peak usage, and provides certified rewrite rules that guarantee non-increasing boundary cost while preserving meaning or bounded approximation error for transformer-scale workloads.
Keywords
- AI
- transformer-scale
- memory wall
- I/O governance
- boundary bytes
- peak hot memory
- admission control
- auditability
- rewrite rules
- KV cache
- activation checkpointing