Originally Posted by Raghar
Then bench it and show the results. I wrote a small Java program just for testing HT, so you have chance to use it. Don't forget to vary RAM speed, because I still don't have numbers on how RAM latency changes the behavior.
It works up to 12 cores. There are instructions http://www.overclock.net/t/1392721/a-small-benchmark-request
From what I seen HT gains are negligible, and considering Tablets and other small platforms are watt limited, the increased power consumption isn't worthy.
Intel reintroduced HT with Nehalem precisely because it added 2%+ of overall performance for every 1% of increase in power consumption, in there tests.
The extra threads don't use any extra power if they aren't in use (indeed, they are parked by default on most consumer OSes), and they should not be used in the minority of scenarios where they do not improve performance or performance per watt.
I'm not able to run your benchmark (possibly because I use JRE6, not 7, since the only things I need Java for stop working correctly with 7), but I do know that threaded memory benchmarks bear almost no resemblance to most real-world workloads.
I am able to test a number of real world programs, and these programs tend to bear out the idea that HT generally helps performance per watt in any scenario you'd want it to be used for in the first place.Edit:
I just did a couple of tests, and on my primary signature system, in Blender I see an ~18% increase in performance with HT for a ~14% increase in power consumption (CPU only measured via the part's ammeter though HWmonitor, which is fairly close to what I see on a clamp ammeter on the EPS12v connector). In 7-zip, I see more than a 38% increase in performance for only 18.3% more power. With TrueCrypt 7.1's AES encryption HT gives me 29% more performance for only 11% more power.
Still doing more tests.Edited by Blameless - 9/9/13 at 8:00am