This is a known bug. The top line stats counts some weird things as rejects so this is inflated. Your miner software counts correct pool responses as rejects, but as they are not really rejects, it cannot associate it to any job which one of GPUs were working on. So it increases the count in the top line, but doesn't increase number of reject in per-GPU stats.

All actually rejected shares are counted in the per-GPU stats and this is the reliable stats which you should believe in.