ResearchBenchmarksLarge-scale model training on Llama 3.3 70B Emulation (train)Large-scale model training on Llama 3.3 70B Emulation (train)Follow15.3Energy Reduction (Iso-Time)Kareus11.45212.45113.4514.449Jan 25, 2026Evaluation resultsMethodMethodLinksEnergy Reduction (Iso-Time)Time Reduction (Iso-Energy)Kareus# Microbatches=64# Microbatches=642026.0115.316.4Kareus# Microbatches=128# Microbatches=1282026.0115.116Kareus# Microbatches=32# Microbatches=322026.0114.116.8Kareus# Microbatches=16# Microbatches=162026.0111.619.1