Oh yeah absolutely, I played with a lot of configurations early on with this problem specifically.
There are pro's and cons to all of them, and in the end it depends on your resources. I favored the cluster-of-clones approach, since SMTP is a service with distribution built in to the protocol. Basically, thats lots of identical servers (Project Gamera) that are entirely self contained. Offloading to another server is an option in that environment too, but again, I was balancing human resource overhead against available hardware. When you run clamd nodes you are incurring costs elsewhere, in terms of the hardware and human(s) maintaining them.
If the idea is to get insane performance (and I mean exponentially higher.. like 1000x or more) on a *single* server, then what you want is this:
http://www.sensorynetworks.com
That is a FPGA (Field Programable Gate Array), specifically designed for open source stuff. Like spamassassin, clamav, snort, etc. If you ever wondered how you could do gigabit speed packet inspection in something that looks like a 1U dell server, thats the secret sauce
Last but not least, there is the distributed object cache model like memcached
http://www.danga.com/memcached/. If you're looking for the swiss-army-knife of solutions, this is a great one. You could equally use this for clustering anti-spam servers, or web sites (this is in fact what facebook, and livejournal use to scale). What I like about this is that its a generic abstraction layer, so you could set up a memcached cluster and use it for multiple applications. Like storing bayes data in spamassassin, and your really big social networking website. All you need to do is adapt the application for the environment, and since its open source already, some folks have been building this into their apps out of the box by default.