Blaming PRNGs sounds like bikeshedding when it's the hot paths in user-land that often do too much or do it badly without respecting the caches, the branch predictor, and/or compiler or don't diagonalize the problem sufficiently to break the hot section of the problem into independent steps to run concurrently and in parallel as much as possible. Some classes of problems have no known parallel solution, and that's okay and beneficial in some cases.
Also, if you feel that strongly, please don't roll your own crypto and use a CSPRNG.
Also, if you feel that strongly, please don't roll your own crypto and use a CSPRNG.