On my i7-860 machines I was seeing a RAC reduction of only about 4-7% when running 4 tasks at a time vs the machines still running 8 at a time. I left HT on, but was just running less tasks.

The problem is that when a real core is made into 2 by HT, they are not fully independent (they share some of the hardware from the original real core). So certain kinds of code cannot process in parallel. So to some extent, how an app behaves on HT vs. real is somewhat app dependent. If you leave HT on, and two HT cores are on the same real core, you may still get the degradation, even if you are running only 4 threads on a HT CPU that can run 8 (or 12) HT threads.

I think you may want to reread how HT works. One thing to note is that each logical core is tied directly to one physical core. You are right that HT is not an advantage for all kinds of code. It has to do with how often the CPU has to toss away an instruction in it's pipeline because of failed branch prediction. The better the branch prediction the less use HT is.
The results I found in testing SETI@home on the i7-860's was that HT does give you more output, but only a small fraction. As those machines are only doing CPU work they will run 8 at a time. If I were using GPU's in them I would leave the logical cores free to feed the GPU's.

You can start BOINC using the affinity command to tell it which cores to use. So if you wanted to tell BOINC to just use the 4 physical cores in the system it is no problem. Here is an example write up I did a few years ago about that which may be helpful.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[