Just to follow up - I have been runnning some tests using cepheus11's tip. The results are dramatically different. These were ran on a 2GB drive in VirtualBox.
Although these tests where ran in a VM,(so a real disk may be quicker in both tests), that's still a difference of over over a factor of 10! Worth thinking about for anyone wishing to securely encrypt a large disk. If these timings extrapolate linearly that means 200GB volume would take about 24 hours by the traditional method and about 2.5 hours by the new - quite amazing.

Yes, encryption algorithms like AES are fast enough to not be the bottleneck. Disc IO is the cryptsetup-method's bottleneck, so you can expect around 30 MB / s. But if someone finds a weakness in AES in the future, it might be possible to tell that this is not (pseudo)random, but just encrypted zero's with a throw-away key.

For a test like you did, you should use the same blocksize for dd in both methods, i.e. 1 MB:

Your second test could have been additionally slower because dd wrote 1-Byte chunks.

1 members found this post helpful.

Yes I can see that to be truly fair that should be done - should've spotted that. I quoted that second test verbatim as this is the way it's usually given in texts on encrypting partitions, I can see how this would speed it up a little.

Anyway, here's the second test with the adjusted byte count set to 1M so they are both operating the same :
Not quite as bad as before but almost. Still a wonderful trick to know about if you are ever encrypting partitions - crypsetup has a lot of options I have never used and I'll be studying it more closely now.