Stagger Lee wrote:

On Fri, 07 Nov 2003 02:08:50 +0000, Martin <ml_news@ddnospamddml1dd.co.uk.dd> wrote:
: Stagger Lee wrote:
: [...]

[...]

I am amazed at your own arrogantly profound ignorance.

: : However, to be brief:
: : 'Hyperthreading' is an Intel marketing term for a very old trick of : keeping the multiple internal CPU functional units simultaneously busy.
: : To the OS, the one physical CPU appears to be two distinct CPUs. Hence, : if only one process is running, the best reported utilisation is 50% (or : rather, 100% for one of the two hyperthread CPUs).
: : You can run a second process to keep the second 'hyperthread' CPU busy.
 
That just isn't right.  I don't know where this non-cited information
was obtained, but my original citation (look it up in the thread
yourself) explains why the *entire* CPU will be devoted to the task at
hand, contrary to your incorrect explanation.  It is an intrinsic part
of the hyperthreading algorithm that the virtual CPU units all
coalesce into one unit when that is the optimal configuration.
Hyperthreading *does not* split the CPU up into some minimum number of
virtual CPUs which is larger than one.

[...]


Insults aside. You are very misinformed.

This non-cited information is elementary Computer Science!


The present multi-million transistor CPUs used in PCs have multiple execution units that can operate in *parallel*.

That is: The CPU is *not* one sequential mechanical entity like a car engine. It is more like an ant colony of interconnected individual units that can each do their own thing.

Intel suffered a P4 design bottle neck whereby with just one process thread, at any one time a lot of the P4's transistors are left idle with nothing to do. An easy performance fix was to use *two* *sets* of *state* hardware and schedule *two* execution threads simultaneously through the one collection of CPU functional units. At any moment, one or more of the functional units will help with one thread. Meanwhile, the remaining functional units can help execute the second thread. Thus, you keep more of all the millions of transistors busy.

(Note: No "coalesce into one unit" can occur, there's no hardware for that! ...Unless you want to play with FPGAs)


If you can be arsed, check it out with the many Intel white papers describing this. You could ofcourse take the trouble to read the Intel P4 CPU specs.

(I consider AMD to have a better design solution (:-). Then again, the Israeli design team's Pentium-Mobile is an excellent high performance design compromise...)


Perhaps you need to correct your incorrect web page.

Regards,
Martin