ResearchBenchmarksOptimizer state memory measurement on Qwen3-235B-A22B (train)Follow872.4Average Optimizer State Memory (MB)PowerStep636.5762,228.3883,820.25,412.012May 11, 2026Evaluation ResultsMethodMethodLinksAverage Optimizer State Memory (MB)PowerStepPrecision=int8, Number...Precision=int8, Number of NPUs=2562026.05872.4AdamWPrecision=fp32, Number...Precision=fp32, Number of NPUs=2562026.056,768