This bug is not specific to amd64; it also occurs in i386 (on a PIII running testing): 2.6.24 does not work, whereas 2.6.22 does.
The printserver in my case is a sweex USB printserver.
Some further observations:
- strace shows that 32k data is sent to the printserver (in fact, the printer starts loading paper) and, then, the sending (CUPS) process stops
- the same problem occurs with the ipp:// protocol instead of lpd://
Fact is that these devices have very limited memory; so my guess is that the problem is related to flow control