These things vary from system to system, from OS to OS, and from subproject to subprojects; only way to know for sure is to try it out for yourself.

That being said, there typically isn't that great of a difference IIRC.

All things being equal, if I assign a number of tasks only to the number of physical cores but turn on hyper threading for any kind of "overhead" would you generally expect better performance?

My observations on Windows:

If you run single thread LLR tasks, one per core, HT on is ~10% worse than HT off. Windows scheduler is doing something bad here, which I nicknames "core smear" as it moves the threads around and I think it overlaps them at times, hence the performance loss. Can be worked around by manually setting affinity to the boinc process, which gets inherited by tasks it runs subsequently.

If you run multi-thread LLR tasks, one thread per core, HT on is essentially no different than HT off. Something about the MT operation means Windows is less stupid.

If the system is a pure cruncher, turning off HT makes things easier to manage. I don't normally run GPU tasks at the same time so don't know how that changes things.

If you run multi-socket NUMA systems with Windows, I'd go HT off unless you like micro-managing where the tasks end up.

My observations on Windows:

If you run single thread LLR tasks, one per core, HT on is ~10% worse than HT off. Windows scheduler is doing something bad here, which I nicknames "core smear" as it moves the threads around and I think it overlaps them at times, hence the performance loss. Can be worked around by manually setting affinity to the boinc process, which gets inherited by tasks it runs subsequently.

Does the program Project Lasso assist in keeping them on the real cpu cores and away from the HT ones?

Does the program Project Lasso assist in keeping them on the real cpu cores and away from the HT ones?

As I understand it they're the same - with HT on each physical core has 2 virtual cores.

(From memory, it's been a while since I checked) windows does something weird like using 7/8 "real" cores when running 8 single threaded tasks on a 16 virtual core system.

I recently tested this with the last challenge and the was no difference running the GPU using the spare threads in run times. ie - Running 16 cores on my 3950X (HT off) then crunching GFN-18 with the GPU, run times were the same as if I was crunching no CPU work and giving the GPU full cores.