Article about Hyperthreading in iX (Message 1 to 3 of 3)
From: Theo Gottwald *
Subject: Article about Hyperthreading in iX
Date: 19 Sep 2002 03:07:10
Message: <3d89779e$1@news.povray.org>
There is an intresting article about Intel's new "Hyperthreading" in the
actual iX Magazin. (heise-Verlag).
In short it says that you can only get a speed improvement if the two
threads do not use the same processor-
execution units (e.g. the floating-point-unit / or the ALU etc.
If thats the case there are even speed-reduces cause the difrent threads
block themselves from memory access.
--Theo
----------------------------------------------------------------------------
---------------------------------Distributed Network-Rendering or Local
SMP-Rendering on all CPU's you have.
With SMPOV und POV-Ray 3.5. Download free at:
http://www.it-berater.org/smpov.htm
From: Le Forgeron
Subject: Re: Article about Hyperthreading in iX
Date: 19 Sep 2002 08:48:51
Message: <3D89C7BC.6010909@free.fr>
Theo Gottwald * wrote:
> There is an intresting article about Intel's new "Hyperthreading" in the> actual iX Magazin. (heise-Verlag).> > In short it says that you can only get a speed improvement if the two> threads do not use the same processor-> execution units (e.g. the floating-point-unit / or the ALU etc.> If thats the case there are even speed-reduces cause the difrent threads> block themselves from memory access.
So, no need for a multi-threaded Pov-engine for the hyperthreading Intel
then.
Only candidates to remain are real multi-processors mobo...
Forecast: Hyperthreading might be a hype if successful, but no real
performance for us...
From: Theo Gottwald *
Subject: Re: Article about Hyperthreading in iX
Date: 21 Sep 2002 00:50:05
Message: <3d8bfa7d$1@news.povray.org>
The article says at another place that the "overall of the execution units"
in the P-IV
are UNUSED for 75% of the time. And that is the base idea behind HT.
Exampel:
"Pipeline stall" when doing unpredicted branches, Integer units unused while
there is FP-OP in Progress etc.
So I would say two things:
1. the speed increase can be 25% to max. of 75% IF the code is specially
done for a hyperthreaded system,
done to "keep all execution units busy". Exactly that: Keep ALL units
busy. Means: intermix FP & ALU
code IN diffrent threads.
2. Normal "multithreaded code can even slow down the program execution
significantly by blocking the threads from memory/by clearing cache lines
(up to -25% !)..
3. It would be better to have a "real" SMP-Core with two physical processors
so you could really get a speed improvement when usinf two FPU-intensive
threads.
4. Btw. that was my conclusion from the iX article. It couldn't verify it.
So lets wait with the last conclusion till we see the first real "nothwood
HT"-Processors. Should be soon the case ....
--Theo
----------------------------------------------------------------------------
---------------------------------
Distributed Network-Rendering or Local SMP-Rendering on all CPU's you have.
With SMPOV und POV-Ray 3.5. Download free at:
http://www.it-berater.org/smpov.htm
