About multi-GPU jobs in real production cluster trace only account for less than 8%

[hjiema]

Hello~
In your paper, you mentioned that "Considering that multi-GPU jobs in real production cluster trace only account for less than 8%[6]". However, upon reviewing the cited paper, I was unable to locate the specific data that supports this statement. I only find "Whereas distributed training is now commonplace, the majority of Philly workloads (> 82%) ran on a single GPU instance when the trace was collected in 2017." Could you kindly clarify how this particular statistic was derived?
Best regards!