## TL;DR
Green AI addresses the growing carbon footprint of machine learning training. From hardware-aware algorithms to sparse training and model distillation, efficiency research is both an economic and environmental imperative as models scale to trillion parameters.

## Core Explanation
Carbon footprint components: (1) hardware manufacturing emissions (embodied carbon); (2) training electricity consumption (operational carbon); (3) inference serving costs (deployment dominates for popular models). Key tools: ML CO2 Impact (Lacoste et al.), CodeCarbon, and carbontracker for real-time emission monitoring.

## Detailed Analysis
Efficiency strategies: (1) Sparse training — train only a fraction of weights (Lottery Ticket Hypothesis, Rigged Lottery); (2) Mixed precision (FP16, BF16, FP8) — reduces compute by 2-4x; (3) once-for-all networks — train one supernet, extract efficient sub-networks; (4) neural architecture search (NAS) for efficiency; (5) datacenter optimization — location with green energy grids. Efficient inference: FlashAttention (Dao et al., 2022) achieves 7.6x speedup through I/O-aware attention.

## Further Reading
- ML CO2 Impact Calculator
- FlashAttention Paper and Triton Implementation
- IEEE Green AI Workshop