TCP, Offload engines and my favorite silver bullet

The recent hot discussion of the TCP offload engines in the context of
potentially reduced reliability prompted me to try and apply one of my
favorite silver bullets to the said problem.

Could addition of standard reliable datagram service (or two services
- connectionless and connection-oriented) to the TCP/IP protocol suite
allow construction of the offload engines with reliability,
undistinguishable from SW solutions?

I believe that for the modern hardware the calculation of the checksum
doesn't constitute significant burden. The major pain in the TCP
processing is segmentation and reassembling (SAR). I am thinking about
reliable datagram protocol that features both segment check sequences
(SCS) and datagram check sequences (DCS).
As a possible implementation the SAR processing, including SCS check
could be effectively offloaded while the final DCS check left to the
software, probably even at the user level to make it as close as
possible to the final destination.

I see only one serious problem with such protocol - the transmitter
would require rather huge memory, at least as far as the current
combination of relatively high bandwidth with poor and unpredictable
round-trip latency remains prevalent.