Repozytorium publikacji - Politechnika Gdańska

In this paper we investigate performance-energy optimization of tokenizer algorithm training using power capping. We focus on parallel, multi-threaded implementations of Byte Pair Encoding (BPE), Unigram, WordPiece, and WordLevel run on two systems with different multi-core CPUs: Intel Xeon 6130 and desktop Intel i7-13700K. We analyze execution times and energy consumption for various numbers of threads and various power caps and demonstrate that energy consumption can be minimized for both CPUs, while metrics such as EDP and EDS could be optimized for the i7-13700K CPU. We further show that percentage energy gain versus execution time loss could be optimized by 3–6% and 7–13%, depending on the algorithm, for the two CPUs respectively, by applying proper non-default power caps.

Autorzy

• mgr inż. Oksana Diakun link otwiera się w nowej karcie ,
• mgr inż. Jan Dobrosolski link otwiera się w nowej karcie ,
• dr hab. inż. Paweł Czarnul link otwiera się w nowej karcie