Analysis of information sources in references of the Wikipedia article "ChatGPT" in French language version.
« We vary tensor parallelism (TP) and pipeline parallelism (PP) in {1, 2, 4} using CodeLlama-34B on a 4×A100 cluster with NVLink, resulting in 9 TP–PP configurations. Average GPU power range from 213.2 W to 355.3 W, peaking at TP=2, PP=1 and dropping with higher parallelism. Energy usage vary between 0.16–0.56 kWh, with the most efficient setups (TP=2, PP=1 and TP=1, PP=2) balancing runtime and power draw. These results suggest that energy efficiency is more strongly influenced by reduced runtime than by minimizing power draw. »