[CentOS] How to select a motherboard -- CPU architectures and chipsets

Thu Dec 15 16:40:50 UTC 2005
Bryan J. Smith <thebs413 at earthlink.net>

Feizhou <feizhou at graffiti.net> wrote:
> Yeah, read about it quite a while ago when it knocked out
> the P4 2.8 processor.

Intel is returning to the architecture.  The original i686
was very efficient, it just didn't scale beyond 1.0GHz
initially -- 1.5GHz with some mid-release asynchronous hacks,
now possible 2.5GHz with more timing/async changes.

Intel hasn't do a full i686 redesign since the original 1994
Pentium Pro.  Understand that's because Intel thought
everything would be IA-64 now.  You have to plan
architectures 5+ years in advance, because it takes at least
12 months to design, and another 24-36 months of timing
closure.

So when EPIC/Predication didn't deliver in the first IA-64
Itanium-Merced, even for IA-64's native ISA (not even looking
at x86 comaptibility), that's when the Netburst team was
formed.  They did the design in only 18 months -- by
completely by-passing timing closure, using very long pipes
that have many stages doing absolutely nothing.

That's where HyperTransport comes in -- it offers two
schedulers in the hope that two virtualized cores can put
more stages to use.  It only works on the horribly
inefficient Netburst architctures -- you will _never_ see
HyperTransport on the Pentium-M or Intel's newer processors. 
The concept of multi-threading on a single core lives and
dies with NetBurst.

The next evolution is multi-threading across multi-core.

> Thanks to their providing docs.

AMD, Intel and nVidia are pretty open with specifications,
except when legally bound otherwise.  E.g., the biggest and
ironic reason why nVidia couldn't share its AGPgart interface
until about 18 months ago was a NDA with Intel (long story).

> Yes, these are the chums in use in the newer boxes I used
> to admin. I loved the 3ware + riser card fiasco though.

Well, when you're pushing 200+ traces at 66MHz, there tends
to be EMF/EMI issues.  3Ware isn't the only one that has had
issues with traces.  Remember, 3Ware 7000/8000/9000S (not
9550SX) use 0 wait state, 64-bit ASIC+SRAM devices.  Trace
length and timing is everything, and heavily affected by
EMF/EMI.

Again, I refer back to the i865 v. i875 issues.  The traces
of a PCB designed for the i875 -- such as the Asus P4C800 --
didn't necessarly work for the exact same chip in the i865,
because it tested to lower tolerances.

> The problems I have are related to their hardware, not
> whether there are good drivers or not.

Actually, the firmware has always been the issue.  The
ASIC+SRAM design was always sound.  They've done some stupid
things, like the RAID-5 firmware update for the 6000 series
(which was _never_ designed for RAID-5).  But other than
that, it's always been a 

> The poor latencies just won't let me use a Pinnacle 
> DC10 board without crashing.

???  Let me guess, RAID-5 on a 9500S?  ;->

> Yeah, whatever. Tyan come out with a board for servers
> based on a VIA chipset for Pentium III cpus and so I got
> to deal with them.

The Tyan "Tiger" series is _not_ a workstation/server
platform, it's the _desktop_ platform.  That's a very common
misnomer.  The "Thunder" is the workstation/server series. 
;->

> I cannot wait for a promise by a Nvidia rep about their
> future chipsets using SATA NCP technology that will allow
> an open source driver to be written to be acted on.

Do you mean NCQ?

Understand that nVidia is _totally_open_ with their designs
right now, including the SATA.  But the SATA hardware just
doesn't do NCQ at all.

But yes, nVidia has been extremely open.


-- 
Bryan J. Smith                | Sent from Yahoo Mail
mailto:b.j.smith at ieee.org     |  (please excuse any
http://thebs413.blogspot.com/ |   missing headers)