> comparing GPU to naive implementations (and sometimes, implementations written in dynamic languages) instead of to highly tuned BLAS or MKL implementations

In your experience, are these comparisons using naive GPU implementations as well? If so, that still seems like a valuable comparison.