Essentially, it's a device-mapper layer that sits in between a backing store (i.e. regular block device) and a virtual DM device.
In most cases, it acts as a virtual RAM disk of the same size as the device. Writes automagically happen to the RAM disk and it flushing stuff out to disk as massive (albeit unsafe) linear writes. The idea is that your system is embedded and battery backed up. When line power goes, you go into an emergency mode where everything gets flushed to disk.
I've been scheming about something like:
AoE <--> DRBD <--> RamBack <--> AoE SAN (staggered across physical shelves)
This would be cool to have ultra-fast but still reliable storage for things like MySQL transaction logs, BDB transaction logs, JFS/Ext3/ReiserFS external journals, and the like. Rather than being backed up by batteries (to provide RamBack with assurance of runtime), DRBD would provide the redundancy against single failure (although separate power feeds would be nice).
When it's SMP-ready I might take a stab at shoving it in production.
UPDATE: I meant DRBD, not GNBD. Also, I recently discovered Write-Intent Bitmaps for MD devices, which also would be great for this kind of backing store.