On May 4, 2007, at 16:20 , Jon Hylands wrote:

> On Fri, 4 May 2007 15:58:40 -0400, Bert Freudenberg  
> <bert at freudenbergs.de>
> wrote:
>
>> I'm pretty sure this is not true, but more like what Damien wrote.
>> Basically all that HT does is having a second register set so you can
>> hold the state of two processes in the CPU at the same time.
>> Execution units are shared. Since one process seldomly utilizes all
>> execution units (due to "pipeline bubbles") the other process can use
>> the free units. This simply enhances throughput, but it's not like in
>> one process you only ever get half the speed.
>
> Okay, well perhaps I misunderstood. I know a couple years ago we did a
> bunch of research into this, because we were running 4-processor  
> servers
> with hyperthreading in each processor, and we were running 4 VisualAge
> server images on each machine. By turning off Hyperthreading, we  
> got a very
> significant speed increase (not quite double, but close). With
> hyperthreading on, we only ended up ever using 50% of the total CPU
> available.

Well, then I'd take your experience over Intel's marketing I guess ;)

The idea actually is rather old, IIRC the Alto was 16-ways  
"hyperthreaded".

- Bert -