[QUOTE=Christenson;251425]If the goal is to maximize the long-term throughput, is it better to have more threads doing disparate tasks or more helper cores pushing the tasks through fewer threads at once?[/QUOTE]Best throughput is generally achieved by having each physical core doing its own test.

Cannot deactivate SMT / HT
 

[QUOTE=Uncwilly;251426]Best throughput is generally achieved by having each physical core doing its own test.[/QUOTE]
My notebook's BIOS doesn't offer disabling hyperthreading so I'm forced to run 1 test per logical core. It's an Intel 2630QM (4 physical, 8 logical).
I assume best throughput is still achieved using no helper threads? I haven't been able to compare both configurations yet.

[QUOTE=Brain;251444]My notebook's BIOS doesn't offer disabling hyperthreading so I'm forced to run 1 test per logical core. It's an Intel 2630QM (4 physical, 8 logical).
I assume best throughput is still achieved using no helper threads? I haven't been able to compare both configurations yet.[/QUOTE]

For non-FFT work, (e.g. TF, sieving, NFS sieving) one worker/instance per logical core is usually best.
For FFT work (e.g. LL, P-1) one worker per physical core is usually best. You could either run 4 workers with each using two threads, (keep the default affinities, which will make each worker use both threads on its own physical core) or 4 workers with each using one thread and the affinities set to 0, 3, 5, 7. You can do this with the AffinityScramble=0357 option in prime.txt (see undoc.txt for more info). AFAIK which of the two is faster can vary and should be experimented on your machine.