Probably the most important tradeoff between nbd/nfs is that NFS by virtue of supporting shared mutation, means client-side caching is either non-existent or barely useful

With NBD you're giving a filesystem driver exclusive access to a block device, the kernel knows the underlying isn't expected to change, so for example no round-trips are required to read or revalidate some cached data, and something like "find /" while slow to run the first time around, will likely be served from cache on a subsequent pass, and thus lightning fast