Green GPU: Integrating Carbon Metrics into GPU Manufacturing with Minimal Disruption

The increasing computational demands of artificial intelligence (AI), especially large language models (LLMs), have brought attention to their environmental impact, particularly operational carbon emissions. However, a significant yet often overlooked factor is the embodied carbon footprint resulting from the manufacturing of AI accelerators like GPUs. This ''hidden'' carbon is crucial and largely neglected in current sustainability efforts for AI hardware. Our analysis reveals that (i) this embodied carbon already constitutes a non-trivial 0.77% of GPT-3's and 2.18% of GPT- 4's operational reported emissions, highlighting a critical blind spot; (ii) while operational carbon of GPT-3 is dominated by compute-intensive kernels such as matrix multiplication and fully-connected layers, the challenge of embodied carbon lies within the complex manufacturing process itself. To effectively green AI, we argue that carbon metrics should be integrated into GPU manufacturing without disrupting existing optimized design and production workflows. We propose (i) linking carbon metrics with performance, power, and area (PPA) optimization metrics to align sustainability with efficiency, utilizing model-based and datadriven approaches; and (ii) using carbon metrics as supplementary constraints within Electronic Design Automation (EDA) tools, rather than disruptive replacements. This approach aims to harmonize sustainability with manufacturing practicality, promoting the development of environmentally responsible AI accelerators without compromising performance, thereby bridging a crucial gap in current green computing initiatives.