Preprint reports faster GB200 training with polynomial function replacements

Researchers report in an arXiv preprint that replacing selected mathematical functions with short polynomial calculations improved complete training-step throughput on GB200 hardware by 2.7%, 2.9%, and 8.0% across three tasks. The replacements approximate functions used inside language models, combining symmetry, rounding to the target number format, and packed arithmetic within the kernels that consume their outputs. A fourth task, sigmoid attention, improved complete-attention forward performance by 7.4% but the complete GPU step by 0.3%. The evaluation included one paired pretraining comparison per task. At common horizons near 100 billion tokens, final smoothed training-loss differences between polynomial and native implementations ranged from -0.107 to +0.079, showing that the substitutions also changed model training behavior.