Comment Re:Dunno (Score 2) 234
They are latency-bound so HT just causes two requests to be outstanding simultaneously, effectively hiding the latency of one request behind another request. UltraSparc T2 was an architecture based around exactly this principle with eight threads per core.
In my experience, if software is well-written (e.g., you make sure you're bandwidth or compute bound), hyperthreads will definitely slow things down. How badly will depend on the exact workload. In all cases enabling hyperthreads will cause your system to behave a lot more variably, which is another thing one may care about.
For the above reasons HT off has been my default setting since it came out-- though ideally it's controlled from software: if you know you're gonna be latency-bound, then use hyperthreads-- otherwise only pin one thread to any single core. But that's without considering Spectre et al...