I got worse results than the paper, 5 percentage points avg

Hello author, may I ask if you used 8 cards or the RTX 8000 GPU single card mentioned in the paper when training MSD8? I used all the configurations exactly the same as decoder only. The 5 fold cross validation results obtained from L20 single card training were on average 5 percentage points worse than the data in the paper. Can you answer this question?