Disclaimer: these are guesses extrapolated from unix.

One thing to check, is dump short-circuiting the write?

OS level writes to /dev/null can just validate their args with the MMU then return true, a dedicated low level backup tool might not have that optimisation. I tend to pipe through wc(1) when benchmarking to avoid getting fooled by that optimization.

On the other hand, high level user space system backup tools focused on writing to tape tend to use multi process tricks to double buffer so as to keep the tape streaming.