On Mon, 2005-12-05 at 17:58 +0200, Oded Arbel wrote:
> On Monday, 5 בDecember 2005 17:05, Gilboa Davara wrote:
> > Either way, Intel is dumping Hyper-threading in future CPUs. In the
> > end, HT was nothing more then a marketing ploy designed to save a bad
> > core design. (The "I've got too many stalled pipelines" P4)
>
> I'm not sure why you say that - I haven't heard of any future intel
> plans to drop HT, OTOH intel is bringing HT to the Itanium line
> ("tukwial" or whatever they name it today), Sun has supposedly "4 way"
> hyperthreading in their Niagara chips, and there are some rumors
> (nothing I can verify) on AMD bringing out hyperthreaded Opterons in
> 2007.
You seem to confuse multiple-cores and DSMT * (or some sort) with
Hyper-threading.
By design, Hyper-threading is a very interesting idea to increase each
core's IPC (by utilization stalled pipe-lines [cache-miss] in-favor of
another "thread") However, due to the problematic design of the P4
itself (very limited bandwidth, small L1, very limited branch history
and register files, etc) and the fact that Hyper-threading was added as
an after-thought to the P4 in-order to improve its lack-luster IPC the
end-result was less then impressive. (Googling around seems to suggest a
+10%/-10% performance diff... nothing to write home about)
The idea of having multiple logical cores in-order to improve
performance is nothing new. Intel's implementation just sucks.
Oh... and all future Intel CPUs will be built around the Pentium M (Not
the P4M) family (or actually, the Yonah/Dotan and the future Merom CPUs)
which doesn't have Hyper-threading support.
http://www.theregister.co.uk/2005/08/23/intel_next_gen_architecture/
As for AMD, can you please post a link that states the above?
> Regarding the kernel's hyperthreading support - after reading the LWN
> article, I'm still confused about the some things:
> *) did the hyperthreading patch made it into the kernel ? for which
> version ?
In general, Hyper-threading looks like a normal SMP machine, so any SMP
capable kernel worked just fine. (I first used it with 2.4.18 and it
worked - OK)
Newer kernel (2.6.6?) added specific schedulers support and optimization
for Hyperthreading.
> *) IIUC, a hyperthreaded CPU will have better performance if the two
> processes it runs "concurrently" share the memory space, as then its
> much less likely to get cache thrashing. OTOH when the two processes do
> not share the same memory, then you will definitely get cache
> thrashing. Is that correct ?
>
* DSMT stands for Dynamic Simultaneous Multithreading.
=================================================================
To unsubscribe, send mail to [EMAIL PROTECTED] with
the word "unsubscribe" in the message body, e.g., run the command
echo unsubscribe | mail [EMAIL PROTECTED]
