The same thing keeps happening to me when viewing a scanned PDF file containing OCR-generated text in paperpile (Viewer with annotations). Deactivating hardware acceleration in Chrome sure helped, but there is still a peak in CPU usage and the viewer is still much slower than with non-scanned pdf files.
I noticed that, when opening problematic (slow rendering) pdfs and highlighting a passage of text, each word seems to be marked separately, which is different from the viewer's behaviour when opening non-problematic (fast rendering) pdfs. Maybe the issue is connected to the way text in documents with OCR is rendered?