CarbonScaling: Extending Neural Scaling Laws for Carbon Footprint in Large Language Models

In three linesCarbonScaling is a hardware-aware analytical framework modeling carbon footprint of frontier LLM training. It integrates neural scaling laws, distributed training strategies, accelerator modeling, and operational/embodied carbon accounting. Source code released on GitHub.

Summary generated by Claude — human-verified