opinions on motherboards for AMD Opteron

"Rui Pedro Mendes Salgueiro" <rps RemoveThis @rena.mat.uc.pt> wrote in message
news:bmm3b7$g28$1@rena.mat.uc.pt...
 > Hello.
 >
 > I am thinking of buying a PC with one AMD 64 bits processor
 > (bi-processor Tyan motherboards cost about $500, so that will
 > wait for next year's budget). I will use Suse Linux 9.0 64 bit.
 >
 > I am not looking for maximum performance, but for reliability.
 > So, I wanted some suggestions / opinions about motherboards.
 >

Have you considederd MSI's dual cpu board targetting workstation usage. Not
much more expensive than the Asus single chip unit.

It is called;

K8T Master2-FAR

Now it is based on the VIA chipset, but before you shudder, consider that
generally people are having more issues with the Nvidia K8 chipset than the
VIA one. Both seem to report very FEW problems anyways.

At least with this MSI board you can throw in an extra CPU when the time is
right. The Tyan and Asus boards are nice, but too damn close to the price of
this dual monster not to think about it.

Tim<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: opinions on motherboards for AMD Opteron 

"Tim Cunnings" <timken> wrote:
 > "Rui Pedro Mendes Salgueiro" <rps DeleteThis @rena.mat.uc.pt> wrote in message
 > news:bmm3b7$g28$1@rena.mat.uc.pt...
  >> I am thinking of buying a PC with one AMD 64 bits processor
  >> (bi-processor Tyan motherboards cost about $500, so that will
  >> wait for next year's budget). I will use Suse Linux 9.0 64 bit.

  >> I am not looking for maximum performance, but for reliability.
  >> So, I wanted some suggestions / opinions about motherboards.

 > Have you considederd MSI's dual cpu board targetting workstation usage.

I had not thought about that one, but one of the local vendors
mentioned it when I talked with him.

 > Not much more expensive than the Asus single chip unit.
 > It is called;
 > K8T Master2-FAR

<a style='text-decoration: underline;' href="http://www.msi.com.tw/program/products/server/svr/pro_svr_detail.php?UID=484" target="_blank">http://www.msi.com.tw/program/products/server/svr/pro_svr_detail.php?UID=484</a>

This is one of those crippled boards in which only one processor
has direct access to the memory (only 4 DIMM slots). This makes
its price a bit less attractive than it seems at first look. Tyan
has a similar model (Tiger K8W (S2875)), but for some reason I
can't find a price for it in pricewatch (not yet released ?).

Looking at MSI's product list,
<a style='text-decoration: underline;' href="http://www.msi.com.tw/program/products/server/svr/pro_svr_list.php" target="_blank">http://www.msi.com.tw/program/products/server/svr/pro_svr_list.php</a>

I see that they have another model: the MSI K8D Master Series
<a style='text-decoration: underline;' href="http://www.msi.com.tw/program/products/server/svr/pro_svr_detail.php?UID=441" target="_blank">http://www.msi.com.tw/program/products/server/svr/pro_svr_detail.php?UID=441</a>

But that one has a price similar to the comparable Tyans (around
$450 at pricewatch).

 > Now it is based on the VIA chipset,

Shudder

 > but before you shudder, consider that generally people are having
 > more issues with the Nvidia K8 chipset than the VIA one. Both seem
 > to report very FEW problems anyways.

At this time I am a bit allergic to VIA chipsets. We have 16
computers which have Gigabyte GA-7ZX motherboads (chipset VIA Apollo
KT133A) and they have been quite troublesome. The problems might
be caused by something else*, but IIRC some of those boards were
replaced under warranty, which in some of the cases solved the
problem. So, I am not willing to try VIA again at this time.

* For instance, the cases are quite small, and we have been exchanging
them when the power supplies or the CPU fans** fail (and both have
been failing)). I also don't trust too much the ethernet boards
(Netgear) and I am starting to replace them.

** because the big fans don't fit on the small cases. The power supply
gets in the way (mini-tower cases: the power supply is on the side of
the motherboard instead of above it as in the medium-sized cases).

 > At least with this MSI board you can throw in an extra CPU when
 > the time is right.

How does this work with the AMDs Opterons ? IIRC, with the Intel
P3s it was reccomended/mandatory that both CPUs were the same
revision. So buying a 2nd CPU later was not guaranteed to work.

 > The Tyan and Asus boards are nice, but too damn close to the
 > price of this dual monster not to think about it.

At this point, all Opteron boards are a bit expensive. Still,
if that means a better quality control than the cheap boards
I don't mind too much at this time.

Thanks for your reply.

--
<a style='text-decoration: underline;' href="http://www.mat.uc.pt/~rps/" target="_blank">http://www.mat.uc.pt/~rps/</a>

..pt is Portugal| `Whom the gods love die young'-Menander (342-292 BC)
Europe | Villeneuve 50-82, Toivonen 56-86, Senna 60-94<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: opinions on motherboards for AMD Opteron 

On Sun, 19 Oct 2003 19:06:09 +0100, Rui Pedro Mendes Salgueiro wrote:

<snip>
 > This is one of those crippled boards in which only one processor
 > has direct access to the memory (only 4 DIMM slots). This makes
 > its price a bit less attractive than it seems at first look. Tyan
 > has a similar model (Tiger K8W (S2875)), but for some reason I
 > can't find a price for it in pricewatch (not yet released ?).
<snip>

Do you have any data to back-up "crippled", or are you just shooting from
the hip?

From what little test data I've seen on this board compared to a board who
does do the memory the "right" way is that it doesn't seem to make a whole
lotta of difference. Of course, the web sites that usually end-up with
this kind of nice hardware to test seem to benchmark the most trivial
things.

Also, don't forget the form-factor issue, the direct-access boards all seem
to be huge.

--
The three questions of greatest concern are -- 1. Is it attractive?
2. Is it amusing? 3. Does it know its place?
   -- Fran Lebowitz, "Metropolitan Life"<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: opinions on motherboards for AMD Opteron 

Douglas Bollinger <dcb.RemoveThis@pa.nospam.net> wrote:
 > On Sun, 19 Oct 2003 19:06:09 +0100, Rui Pedro Mendes Salgueiro wrote:
 >
 > <snip>
  >> This is one of those crippled boards in which only one processor
  >> has direct access to the memory (only 4 DIMM slots). This makes
  >> its price a bit less attractive than it seems at first look. Tyan
  >> has a similar model (Tiger K8W (S2875)), but for some reason I
  >> can't find a price for it in pricewatch (not yet released ?).
 > <snip>
 >
 > Do you have any data to back-up "crippled", or are you just shooting from
 > the hip?

The opteron supports a seperate 5.3 GB/sec (peak not observer) memory
bus PER cpu. So boards sold in the 4+0 configuration are basically
ommiting the second memory bus. While benchmarks often only use a single
memory bus and don't notice, this can seriously impede performance.

 > From what little test data I've seen on this board compared to a board who
 > does do the memory the "right" way is that it doesn't seem to make a whole
 > lotta of difference. Of course, the web sites that usually end-up with
 > this kind of nice hardware to test seem to benchmark the most trivial
 > things.

Agreed.

 > Also, don't forget the form-factor issue, the direct-access boards all seem
 > to be huge.

Indeed dual 940 pin sockets, dual sets of 4 dimms, AGP, and a couple PCI
add up to a pretty decently sized board.


--
Bill Broadley
Mathematics
UC Davis<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: opinions on motherboards for AMD Opteron 

Douglas Bollinger <dcb.TakeThisOut@pa.nospam.net> wrote:
 > On Sun, 19 Oct 2003 19:06:09 +0100, Rui Pedro Mendes Salgueiro wrote:

 > <snip>
  >> This [MSI K8T Master2-FAR] is one of those crippled boards in
  >> which only one processor
  >> has direct access to the memory (only 4 DIMM slots). This makes
  >> its price a bit less attractive than it seems at first look. Tyan
  >> has a similar model (Tiger K8W (S2875)), but for some reason I
  >> can't find a price for it in pricewatch (not yet released ?).
 > <snip>

 > Do you have any data to back-up "crippled",

Not really, but it seems quite obvious that forcing the second
processor to access memory through the first (so, extra latency)
is crippling it, but I have not searched for benchmarks. I would
be interested in seeing SPECrate results, but I kind of doubt that
vendors (AMD or others) have benchmarked the "crippled" boards.

I am now looking at the results at:
  http://www.spec.org/cpu2000/results/rint2000.html
  http://www.spec.org/cpu2000/results/rfp2000.html

AMD has a number of results published. Selecting all the 5 that test
Einux A4800, all 5 have 8x512MB PC2700 DDR ECC Registered SDRAM CL2.5,
so I think they are using both banks (each bank has only 4 slots, right?).

Has C'T published something ? I searched a bit:
<a style='text-decoration: underline;' href="http://www.google.pt/search?q=site%3Aheise.de+opteron+SPECint_rate_base2000" target="_blank">http://www.google.pt/search?q=site%3Aheise.de+opteron+SPECint_rate_base2000</a>

<a style='text-decoration: underline;' href="http://www.heise.de/ct/Redaktion/as/spec/ct032098/" target="_blank">http://www.heise.de/ct/Redaktion/as/spec/ct032098/</a>
<a style='text-decoration: underline;' href="http://www.heise.de/ct/Redaktion/as/spec/ct0309106/" target="_blank">http://www.heise.de/ct/Redaktion/as/spec/ct0309106/</a>

Looking at 2 results, that seem comparable:
<a style='text-decoration: underline;' href="http://www.heise.de/ct/Redaktion/as/spec/ct0309106/newisys/CFP2000.014.html" target="_blank">http://www.heise.de/ct/Redaktion/as/spec/ct0309106/newisys/CFP2000.014.html</a>
<a style='text-decoration: underline;' href="http://www.heise.de/ct/Redaktion/as/spec/ct032098/msi_9130/2cpu/CFP2000.004.html" target="_blank">http://www.heise.de/ct/Redaktion/as/spec/ct032098/msi_9130/2cpu/CFP2000.004.html</a>

The values obtained are quite near (18.6 for the MSI K8T-Master2
(the board which was being discussed) , with 2x1024MB PC2700 Reg-ECC
DDR SDRAM 2533 versus 18.4 for the Newisys with 6 GByte DDR333,
presumably in 2 banks), despite the MSI using AMD Opteron 246
(2000 MHz), while the Newisys uses AMD Opteron 244 (0F50) (1768 MHz) *.
So it seems that the extra memory bank makes a difference.

* the page says "CPU(s) enabled: 1" but this must be a typo,
because the runtimes for two copies of the benchmarks are quite
similar to runtimes for the MSI board, which is said to be using
2 cpus.

A benchmark where the 2 banks should have a big effect would be to run
2 copies of STREAM at the same time (assuming that the OS manages to
put each instance in one CPU and use the local memory, it should be
possible to get near 100% of the bandwidth in both). I found a report
of such a test (in ppt format):

<a style='text-decoration: underline;' href="http://wwwseminars.web.cern.ch/wwwseminars/2003/2003-OtherFormats/t-20030903.ppt" target="_blank">http://wwwseminars.web.cern.ch/wwwseminars/2003/2003-OtherFormats/t-20030903.ppt</a>

In slide 13 there is:
    1x Stream: 2x Stream: 4x Stream:
2x Opteron, 1.8 GHz,
HyperTransport: 1006 1671 MB/s 975 1178 MB/s 924 1133 MB/s
2x Xeon, 2.4 GHz,
400 MHz FSB: 1202 1404 MB/s 561 785 MB/s 365 753 MB/s

It is not very obvious what is being reported here, because the
Opteron numbers seem too good. It is understable that the Xeon
numbers are halved going from 1 STREAM to 2 and then halved again
with 4 (shared bus). It is also understable that the Opteron numbers
don't fall much from 1 STREAM to 2. It is less understable that 4
STREAMs are almost as fast than 2, since there are only 2 memory
buses.

I have just run STREAM on my computer and I saw this message:

"Each test is run 100 times, but only
the *best* time for each is used."

So, unless the benchmark is modified**, if one has two copies running
at the same time, it might happen that one manages to do a run while
the other is inactive for some reason (it has already finished, for
instance), and that is the number which will be reported.

** I have not verified if this is already checked. Ignoring the
few first and last runs should be enough to prevent the problem.

 > Also, don't forget the form-factor issue, the direct-access boards
 > all seem to be huge.

For my purposes, it doesn't matter much. I'll probably use a tower
case, anyway. But such boards are able to fit in a 1U computer.
See the Newisys 2100 on the ppt mentioned above.

--
<a style='text-decoration: underline;' href="http://www.mat.uc.pt/~rps/" target="_blank">http://www.mat.uc.pt/~rps/</a>

..pt is Portugal| `Whom the gods love die young'-Menander (342-292 BC)
Europe | Villeneuve 50-82, Toivonen 56-86, Senna 60-94<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: opinions on motherboards for AMD Opteron 

On 6 Nov 2003 20:05:03 -0000, Rui Pedro Mendes Salgueiro
<rps DeleteThis @rena.mat.uc.pt> wrote:
 >Douglas Bollinger <dcb DeleteThis @pa.nospam.net> wrote:
  >> On Sun, 19 Oct 2003 19:06:09 +0100, Rui Pedro Mendes Salgueiro wrote:
 >
  >> <snip>
   >>> This [MSI K8T Master2-FAR] is one of those crippled boards in
   >>> which only one processor
   >>> has direct access to the memory (only 4 DIMM slots). This makes
   >>> its price a bit less attractive than it seems at first look. Tyan
   >>> has a similar model (Tiger K8W (S2875)), but for some reason I
   >>> can't find a price for it in pricewatch (not yet released ?).
  >> <snip>
 >
  >> Do you have any data to back-up "crippled",
 >
 >Not really, but it seems quite obvious that forcing the second
 >processor to access memory through the first (so, extra latency)
 >is crippling it, but I have not searched for benchmarks. I would

Interesting bit to think about here. Assuming that you OS doesn't do
any fancy tricks with regards to NUMA systems or processor affinity
(as of a few years ago most operating systems didn't, though it's been
added now), the difference isn't much. Reason? If you have no
processor affinity, then on a dual-processor Opteron board with both
memory controllers you essentially have a 50/50 shot of the memory you
need being local/remote. Memory just kind of goes wherever. So each
processor does half of it's memory accesses locally and half remote.
Switch it to a setup where you have memory on only one processor, you
then have one processor doing 100% local memory access and one
processor doing 100% remote memory access. Total amount of local and
remote memory accesses are the same in either case.

Of course, there's a bit more at work here. First off, this assume
that your two processors are doing equal amount of work. If your
software tends to assign more work to processor 0 (the one with local
memory) than to processor 1 (no local memory), then you could actually
see a bit of a benefit to only having memory attached to one chip.
Alternatively if one of your chips does lots of memory access all over
the place and the other is sitting around processing all from cache,
you can see a fairly large benefit from having memory on both chips.
Also, if you have memory on both chips, presumably your memory
requests are split roughly 50/50 in terms of direction, and with
hypertransport having two unidirectional links, you have twice as much
bandwidth than if all your traffic is one-way.

All that being said though, I'd still recommend sticking to dual
Opteron boards that have memory off BOTH processors. Particularly as
operating systems continue to improve, this should give you optimal
performance.

-------------
Tony Hill
hilla <underscore> 20 <at> yahoo <dot> ca<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: opinions on motherboards for AMD Opteron 

I'll warn about the SK8N. I'm having a bear of a time getting Suse
9.0 installed because it won't recognize the onboard Promise RAID
controller. I've gone through their help docs, but so far they
haven't helped.

- Greg

Rui Pedro Mendes Salgueiro <rps RemoveThis @rena.mat.uc.pt> wrote in message news:<bmm3b7$g28$1@rena.mat.uc.pt>...
 > Hello.
 >
 > I am thinking of buying a PC with one AMD 64 bits processor
 > (bi-processor Tyan motherboards cost about $500, so that will
 > wait for next year's budget). I will use Suse Linux 9.0 64 bit.
 >
 > I am not looking for maximum performance, but for reliability.
 > So, I wanted some suggestions / opinions about motherboards.
 >
 > The 2 boards I am considering are:
 >
   > Asus SK8N Tyan K8S (S2850)
 > ----------------------------------------------------------------------------
 > Chipset Nvidia nForce 3 AMD-8111 I/O hub
   > pro 150 Winbond W83627HF Super I/O
   > Winbond W83782D Hardware Monitor
 > ----------------------------------------------------------------------------
 > Advantages - AGP slot - only PCI (32 bits) slots
 > Asus - audio (6 channel) - no audio
   > including digital output
   > - Promise R20378 RAID - optional RAID
   > (1 IDE, 2 Serial ATA) (4 Serial ATA)
   > - firewire IEEE1394 - no firewire
   > - 6 USB 2.0 - 4 USB 1.1
 > ----------------------------------------------------------------------------
 > Advantages - only 10/100 ethernet - Gigabit Ethernet (2 ports)
 > Tyan - no integrated graphics - integrated graphics with
   > separated video memory
   > - maximum IDE disk size - 48 bit LBA (for big disks)
   > 128GB = 137438953472 bytes ?
 > ----------------------------------------------------------------------------
 > Tied or - 4 DDR slots (8GB maximum)
 > nearly tied price $200 $180
 > ----------------------------------------------------------------------------
 >
 > If I was building a workstation, the Asus would look like the better option,
 > despite the lack of Gigabit Ethernet. But for low-end server use, they are
 > quite tied, with some advantage for the Tyan.
 >
 > About the chipsets, is there enough experience yet to know if either
 > one is more troublesome than the other ? And will AMD drop theirs in
 > a few months, or this time they are staying on the chipset business ?
 >
 > Thanks in advance for any replies.<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: opinions on motherboards for AMD Opteron