[NTLUG:Discuss] Re: Chaintec MOBO -- brand name is nothing (model+revision is)

[Bryan J. Smith]

On Sun, 2004-11-28 at 20:11, Dennis Myhand wrote: 
> Is anyone using a chaintec motherboard,

Model number?

> or does anyone have anything good to say about them?

Brand name is pretty much irrelevant these days, unless you're
talking about honoring warranties or other retail/support detail.
Failure of a device has nothing to do with a vendor these days,
although warranty issues are definitely something to attach to
a vendor name.

The model number and _specific_ revision of a device is typically
the only way to tell these days.  Look up the specific model and,
if it is know, the revision.  So many vendor change manufacturers
on-a-whim, sometimes even mid-model.  My favorite example in the
hard drive world is Western Digital, they outsource their manufacture
to different companies continuously -- for the same model.

What is the exact model number (and revision if known)?

> I am soon to be building a new system and I am wondering about
> using lowerr cost motherboard in an AMD Athlon64 system with a
> socket 754 chip.  Thanks, Dennis in Victoria

First thing to understand about A64/Opteron, there is no such thing
as the Front Side Bottleneck (FSB) anymore.  People still try to
say it has a 600-1000MHz FSB, but they are only talking about the
1200-2000 Megatransaction (MT) HyperTransport to I/O (and other
CPUs in the case of multiprocessor).  Memory is local to each CPU
(and even I/O is too, depending on the design).

For more on how AMD's NUMA/HyperTransport is different (largely
written for workstations/servers):  
  http://www.samag.com/documents/sam0411b/  

AMD has 3 sockets:
  Socket-754:  Uniprocessor       1x DDR, 1x HT
  Socket-940:  Uni/Multiprocessor 2x Reg-DDR, 1-3x HT (per CPU)
  Socket-939:  Uniprocessor       2x DDR, 1x HT

Originally there was only going to be Socket-754 and 940.  But the
added cost of Registered DDR caused AMD to rethink the strategy.  A
lower costing Socket-939 was conceived.

Socket-939 (and 940) has 2 _true_ DDR channels, a _true_ 128-bit
path to memory (per CPU in the case of 940).  With PC3200/DDR400, this
is 6.4GBps (per CPU in the case of 940).  Socket-939 also typically
has a 2000MT (8.0GBps) HyperTransport interconnect to I/O.  I know
this gets confusing, but if you just want numbers in comparison,
see this table on "Aggregate Front Side Throughput" (AFST):  
  http://www.samag.com/documents/s=9408/sam0411b/0411b_t1.htm  

In a nutshell, Socket-754 has an maximum AFST of 9.6GBps, Socket-939
has 14.2GBps.  Socket-940 varies by type/number of CPUs.

With that said, what mainboard should you buy?

A cheap Socket-754 is one way to go for well under $100.  Socket-754
Athlon64's are around $120, worth the extra few bucks over a Socket-754
Sempron.  I might consider the Sempron for Windows, because it is 32-bit
only (most of x86-64 is disabled), but for Linux, I'd go the full A64.
One thing to note is that Sempron is rated against the _Celeron_, A64 is
rated against the P4 -- so Sempron is rated higher than its equivalent
A64.

If you can wait, PCI-Express Socket-939 (and 940) mainboards will arrive
in December at start sub-$150.  Add in the slight premium for a Socket-939
A64 3000+ and 2 DIMMs you have one quick, kick-butt mainboard.  Especially
with a low-cost, $125 nVidia GeForce 6600 128MB that really blows away a
far more expensive AGP card.  But PCI-Express is more about segmenting
I/O than just an improvement in graphics data transfer rates (DTRs).

The new nVidia nForce4 chipset, along with others from ViA and SiS, will
provide ATA, NIC and other I/O components on their own peripheral bus.  
This is important as a "shared" 0.125GBps (32b at 33MHz) PCI bus is not
enough to carry disk, network and audio these days.  Many people using
cheap or on-chipset audio get "latency/choppiness" because ATA or NIC
data transfers are hogging the PCI bus.


On Sun, 2004-11-28 at 21:25, Jeff wrote:
> I once made the mistake of buying a chaintech MB.  replaced it under 
> warranty 4 times before giving up and just getting a FIC.  the FIC has 
> been running solid for 2 years now.

Again, brand name is nothing.  ECS, FIC and several other Tawainese
vendor cross-manufacturer each other's products under the 9 (is it?)
company "PCChips" consortium.  Some are branded PCChips, typically the
"left-overs" of either surplus supply, or components that did not test
to adequate tolerances for their regular "retail" products.

There was also the issue where a Tawainese company stole an incomplete
resin design from a Japanese company.  It was used in capacitors in
the mainboards of many vendors, including the larger Tawainese vendors
like ECS.  Again, outside of retail/warranty issues, vendor name means
_nothing_.  Model is everything -- exact revision of the model is the
best thing to know, as some vendors change entire ICs/manufacturers
mid-model.

ECS is so large (much larger than people realize), I had some K7S5As
that ran excellent, and some that were just crap.  Closer inspection
revealed they were clearly manufactured under very different conditions
and different QA considerations.  I quickly noted differences in etchings
on the PCBs, even though the PCBs were exactly the same.  Some looked
like they had been through a high-QA plant, whereas some looked like
the ICs had been attached manually with a solider.  ;->

Here's another, current example.  One mainboard that has major issues with
major kernel 2.6 distros (sometimes select kernel 2.4 distros as well) is
the Asus P4P800 series with the Intel i865 chipset.  The Asus P4C800 is
the _exact_same_ PCB _and_ chipset logic, only the Intel i875 model.  What
is the difference?

The Intel i875 tests to higher tolerances and better driver (current) than
the i865.  The P4C800 was designed and released first, specifically for the
i875.  Asus assumed they could use the same PCB design for the P4P800 with
the cheaper i865 chipset which came out shortly afterwards.  While even the
first P4C800 had 0 issues with Linux, various P4P800s started to have all
sorts of issues -- most of which were hanging when various distro installers
started probing the hardware.

_Exact_ same PCB, _exact_ same logic throughout.  Reality is that the P4P800
had to have some redesigns, and latter series boards seem to be better.

This is just one, very small example.  So it is very important to check the
per-model reviews, especially if you can match up a review to the exact
revision.  Vendor brand name doesn't mean squat.  I've had both crappy and
excellent mainboards from Abit, Asus, Chaintech, ECS, Gigabyte, Microstar
(aka MSI), Tyan, etc...  And it's _rarely_ an issue of chipset, but how
the mainboard uses it.


On Thu, 1970-01-01 at 00:59 (FIX DATE? ;-), Kevin E. Ivey wrote:
> Howdy,
> I have use Chaintech boards since my first VLB board back in 1994.  That 
> is how I came to love SiS chipsets.  Nowadays, Chaintech builds just 
> about anything with a circuit board, but they make some pretty good 
> stuff.  German toughness became Taiwanese cleverness, IMO.
> They seem to use VIA chipsets and NVidia GPU's exclusively, so you 
> probably can't go wrong.

Typically that's a matter of vendor alignment than good design.  In
fact, beware of low-cost vendors who are probably taking the lot of
"QA failed" products and re-testing them to lower tolerances.  Anyone
in the PC Chips consortium and that new HD consortium (can't remember
the name) should _beware_.

> I think the 754 board will become cheap very soon due to the 939-pin 
> Athlon64 becoming the standard.  Great, I finally bought my A64 and they 
> change the form factor.....

They didn't change it at all.  There were _always_ the two designs,
ClawHammer (uniprocessor) and SledgeHammer (uni/multiprocessor).  The
former had exactly one memory and one HT channel, whereas the latter
had two memory and one or more HT channels.

Socket-939 is merely a slight change in Socket-940, removing the
requirement for Registered DDR SDRAM and jacking the HyperTransport link
to 2000MT (8.0GBps).  The Registered DDR SDRAM requirement was really
overkill, and only necessary for when you have 8 DIMMs per CPU (16 or 32
in 2 or 4 way server mainboards).  Under JEDEC standards for PC2100/2700
(DDR266/333), 2 DIMMs per channel (upto 4 per Socket-939 CPU) is well
within spec.

BTW, because AMD NUMA/HyperTransport localizes memory, if you use a
lower synchronous clock, it does _not_ affect any other part of the
system.  This is unlike Front Side Bottleneck (FSB) interconnects like
Intel AGTL+ (Socket-478, 604, LGA-775, etc...) where CPU-memory-I/O is
all synchronous at the _northbridge_ (what Intel calls the "memory
controller hub," MCH).

To make an analogy to Ethernet between CPU, memory and I/O:  
  Intel AGTL+ (all CPUs, from PPro to P4/"Prescott") = 3 port hub
  AMD/Digital EV6 (32-bit Athlon, 64-bit Alpha 21264) = 3-16 port switch
  AMD NUMA/HyperTransport (A64/Opteron) = direct mesh** (node to node**)

**I.e., it's like having every system with direct Ethernet connections
to every other system.  In our case, direct connections between CPUs,
memory and I/O.

Socket-754 has 2.  Socket-939 has 3.  Socket-940 has 3-5 _per_ CPU.


-- 
Bryan J. Smith                                    b.j.smith at ieee.org 
-------------------------------------------------------------------- 
Subtotal Cost of Ownership (SCO) for Windows being less than Linux
Total Cost of Ownership (TCO) assumes experts for the former, costly
retraining for the latter, omitted "software assurance" costs in 
compatible desktop OS/apps for the former, no free/legacy reuse for
latter, and no basic security, patch or downtime comparison at all.