Is it worth it? Opteron 144 vs. Athlon 64 3200+

[Nate Edel]

Looking at costs for a new system, it looks like I can get an ASUS SK8V +
Opteron 144 for about the same cost as an ASUS Socket 754 board and an
Athlon 64 3200+

I'd like to get the Opteron system, mainly for the extra DIMM slot although
the availability of the non-Nvidia chipset is a small plus as well (looking
at driver availability under Linux).

My sense is, though, that as long as I go with the 3200+ and not the 3000+,
the 200mhz is going to make a bigger difference than the second memory
channel, and /that/ is the point I'm curious on.

For what it's worth, of what I do that's performance-relevant, my workload
is mostly large builds with GCC and some simulation under Linux.

I also do a little bit of home video stuff, and a little bit of gaming
(although everything I play right now runs fine on my P4 1.7, and I'm
planning on transferring my Radeon 8500, not replacing it for now.)

So... the real answer is I'll probably just say "whatever" and push the
budget a little to get a Opteron 146 (148s or FXs are just nowhere near in
the cards, budget wise... though a 3400+ or dual 140s might be.)

 

[Grumble]

So... the real answer is I'll probably just say "whatever" and push the
budget a little to get a Opteron 146 (148s or FXs are just nowhere near in
the cards, budget wise... though a 3400+ or dual 140s might be.)

Do you think dual 140s work?

If one believes AMD, one needs 240s i.e. $396 instead of $326.

 

[Rob Stow]

Do you think dual 140s work?

If one believes AMD, one needs 240s i.e. $396 instead of $326.

Assuming you are using yankeebucks, shouldn't those Opty 240
prices be about $100 lower ? I have used exactly two
Opty 240's so far and they were $22x from www.upgradesource.com
in January.

 

[Nate Edel]

Do you think dual 140s work?
If one believes AMD, one needs 240s i.e. $396 instead of $326.

Ooops, no, typo... dual 240s rather.

And what country's dollars/currency are those prices in? Because the prices
on 140s and 240s here in the states are about $180-200 for 140s, and about
$215-225 for 240s.

The cost difference of a dual 240 to a single 146 is about $200 for the
motherboard, and about $140 for the two chips.

 

[krw]

Ooops, no, typo... dual 240s rather.

And what country's dollars/currency are those prices in? Because the prices
on 140s and 240s here in the states are about $180-200 for 140s, and about
$215-225 for 240s.

The cost difference of a dual 240 to a single 146 is about $200 for the
motherboard, and about $140 for the two chips.

The Tyan K8W (S2875) supports dual Opterons and is only $20-$30 more
than the single processor K8WS (S2875S). In fact I was seriously
considering it until I discovered (by looking at a picture of the
board) that only one of the processors has memory attached to it.
That's got to do whacky things to performance.

 

[Nate Edel]

The Tyan K8W (S2875) supports dual Opterons and is only $20-$30 more
than the single processor K8WS (S2875S).

Or about $40 more than an Asus SK8V... still quite inexpensive. Is it still
in production? It does have AGP, which was missing on the first-generation
dual boards, so it can't be _that_ old.

In fact I was seriously considering it until I discovered (by looking at a
picture of the board) that only one of the processors has memory attached
to it. That's got to do whacky things to performance.

Well, not wacky, but definitely will favor one processor a bit over the
other.

If those will run with only one processor, perhaps picking one up with a
single 240 or 242 and getting the second later would actually be economical.

 

[Douglas Bollinger]

The Tyan K8W (S2875) supports dual Opterons and is only $20-$30 more than
the single processor K8WS (S2875S). In fact I was seriously considering
it until I discovered (by looking at a picture of the board) that only one
of the processors has memory attached to it. That's got to do whacky
things to performance.

No. If you bother to actually check out the benchmarks, you will find that
the memory setup on the K8W works just find. Unless you are running a NUMA
aware OS (not windows), that's the way the memory works on the other board
anyway.

 

[Rob Stow]

No. If you bother to actually check out the benchmarks, you will find that
the memory setup on the K8W works just find. Unless you are running a NUMA
aware OS (not windows), that's the way the memory works on the other board
anyway.

Not true. You do *not* need a NUMA-aware OS in order to
take advantage of the way a board like the S2885 has
four DIMM slots for each processor. In the absence of
a NUMA-aware OS that micromanages the RAM allocation,
each processor has a built-in affinity for using the
RAM attached to it before using RAM attached to other
processors. There is a better/more detailed explanation
in some of the many Opteron PDFs downloadable from AMD.

I have used one S2875 and three S2885s so far.
I have compared an S2875 vs S2885 where the systems were
identical in all components except for the motherboard.
In simple tests with XP and the 64 bit Windows beta, the
S2885 was a *minimum* of 8% faster in everything - provided
that I was doing things that stressed both CPUs.

If you are doing something that only stresses one CPU,
then the two motherboards perform comparably. For example,
if you encode a video with TMpegEnc and have nothing else
running but the basic Windows services and drivers.

Change it so that you have something like TMpegEnc encoding
videos while you simultaneously have LAME encoding WAV==>MP3
and suddenly the S2885 outperforms the S2875 by about 21%.

Another major disadvantage of the S2875 is that it
has no PCI-X slots - an unacceptable omission for a workstation
motherboard. If you want to use a good RAID controller
you are therefore screwed.

There are serious downsides to the S2885, however.
- Only one legacy PCI slot, as noted above. There is, however,
a jumper that lets you drop two of the PCI-X slots down to
33 MHz, so you can still use two more 3.3 Volt 32 bit cards.
It took me a while to realize that the jumper existed and I
could hence install both a USB 2.0 card and a sound card.
- EATX form factor: fewer cases to choose from and they cost more.
- EPS12V/SSI power supply required. By comparison, the S2875 can
use either ATX12V or EPS12V. Again, fewer PSUs to choose from
and they cost more.
- Price. An S2885 costs about twice as much as a S2875. By the
time you account for needing a more expensive case and a more
expensive PSU, the actual price difference between an S2885 system
and an S2875 system is about $500 (Canadian), or about $350 US.

Note that the absence of built-in USB 2.0 is annoying with
both the S2885 and the S2875. Users of either motherboard
will probably want a USB 2.0 PCI card.

 

[Tony Hill]

Ooops, no, typo... dual 240s rather.

And what country's dollars/currency are those prices in? Because the prices
on 140s and 240s here in the states are about $180-200 for 140s, and about
$215-225 for 240s.

He's talking about the price for two chips, using AMD's processor
pricing guide list here:

http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_609,00.html?redir=CPT301

Note that these are the official prices that AMD sells to distributors
at in lots of 1000. Retail prices can (and often are) either higher
or lower than these prices depending on a variety of factors.

 

[Nate Edel]

He's talking about the price for two chips,

Ah, gotcha. That's the point I missed.

 

[KR Williams]

No. If you bother to actually check out the benchmarks, you will find that
the memory setup on the K8W works just find. Unless you are running a NUMA
aware OS (not windows), that's the way the memory works on the other board
anyway.

Why would I consider Win? My issue is the latency and bandwidth
of one processor accessing the other for *all* transactions.
This seems rather unseemly. ;-) OTOH, the price is right.

 

[KR Williams]

Or about $40 more than an Asus SK8V... still quite inexpensive. Is it still
in production? It does have AGP, which was missing on the first-generation
dual boards, so it can't be _that_ old.

Indeed it looks like the single processor version isn't
available. The dual is available from reputable sources for $230
or so.

Well, not wacky, but definitely will favor one processor a bit over the
other.

Since one will be severely memory bandwidth and latency
handicapped, I'll stand by my "wacky" assessment. ;-)

If those will run with only one processor,

They will, according to all documentation.

perhaps picking one up with a
single 240 or 242 and getting the second later would actually be economical.

Sorta my thoughts. Though I'm still looking. It's been a *bad*
month, so I may have to wait some.

 

[Tony Hill]

Why would I consider Win? My issue is the latency and bandwidth
of one processor accessing the other for *all* transactions.
This seems rather unseemly. ;-) OTOH, the price is right.

It's not actually as bad as it might seem initially. First off,
picture a system with absolutely no NUMA support at all. In such a
setup with a dual-Opteron board with memory hanging off either CPU (ie
the Tyan Thunder board), you have exactly a 50/50 shot of the memory
being local or remote.

Now compare that to the an Opteron board with memory hanging off only
one CPU (ie the Tyan Tiger board). Here you ALWAYS have local memory
for one chip and always have remote memory for the other chip. Given
an equal workload between the two you end up with 50% of memory
requests being local and 50% being remote, exactly the same as in the
first case.

Of course, there are some advantages to having memory hanging off both
processors. First off it doubles the number of sockets and amount of
memory you can use. Second, any NUMA optimizations at all are a win.
NUMA optimizations are fairly weak in Windows (though they do exist in
WinXP and even more so in Win2K3 Server), and the Linux 2.4 kernel
isn't so hot either. The Linux 2.6 kernel improved on this a bit, but
you're still probably looking at no better than a 60/40 split for
local/remote memory (very rough guesstimate).

The third advantage is that you can potentially get a bit more
bandwidth for any given task by having memory hanging off both
processors, though I don't know that this adds much. The Opteron
doesn't seem super-bandwidth hungry and your memory would have to be
laid out in a bit of an ideal pattern to see the benefit here.

Anyway, end result is that the two-memory controller setup is usually
(though not always) going to be faster than using only a single memory
controller, however the difference often isn't huge. Most tests I've
seen have it at about 5-10% except in extreme cases, and often even
less than that. For single-threaded stuff the single memory
controller can often be faster since all your memory access is always
going to be local as long as your running on the right CPU.

 

[Douglas Bollinger]

Not true. You do *not* need a NUMA-aware OS in order to take advantage of
the way a board like the S2885 has four DIMM slots for each processor. In
the absence of a NUMA-aware OS that micromanages the RAM allocation, each
processor has a built-in affinity for using the RAM attached to it before
using RAM attached to other processors. There is a better/more detailed
explanation in some of the many Opteron PDFs downloadable from AMD.

Cool. That's not what I read on the hardware review sites, but then again
that doesn't mean very much. Usually every site that doesn't get quite the
memory bandwidth or performance they are looking for with the "separate
memory " Opteron systems complain about the lack of a NUMA based OS, but
it's nice to see this is not necessary.

I have used one S2875 and three S2885s so far. I have compared an S2875 vs
S2885 where the systems were identical in all components except for the
motherboard. In simple tests with XP and the 64 bit Windows beta, the
S2885 was a *minimum* of 8% faster in everything - provided that I was
doing things that stressed both CPUs.

If you are doing something that only stresses one CPU, then the two
motherboards perform comparably. For example, if you encode a video with
TMpegEnc and have nothing else running but the basic Windows services and
drivers.

This site has a review of the two different kinds of Opteron boards:

http://www.gamepc.com/labs/print_content.asp?id=dualduel

Of course, they do _very_ little multithreaded tests. That's the trouble
with these sites, they get all the cool hardware but do games tests on
these boards. Who cares? Slap a high-end RAID SCSI or SATA card in and
see if the board locks-up under intense IO. Look for problems, not just
make pretty graphs. I know that the system I'm using now, a AMD MPX based
Gigabyte GA-7DPXDW+ certain has it's shares of foibles that none of the
hardware reviews I've seen ever found.

 

[Rob Stow]

Cool. That's not what I read on the hardware review sites, but then again
that doesn't mean very much. Usually every site that doesn't get quite the
memory bandwidth or performance they are looking for with the "separate
memory " Opteron systems complain about the lack of a NUMA based OS, but
it's nice to see this is not necessary.

And a lot of those sites only review the cheap boards that have
all of the DIMM slots attached to CPU0, instead of having separate
DIMMs for each CPU.

This site has a review of the two different kinds of Opteron boards:

http://www.gamepc.com/labs/print_content.asp?id=dualduel

Of course, they do _very_ little multithreaded tests.

The kinds of tests they did still could have been a lot more revealing
if they had done them simultaneously instead of one at a time. Eg.,
encode a video while testing a game instead of doing them separately.

What the heck is the point of testing a dualie if the tests don't
stress both of the CPUs at the same time ?

And if the tests don't stress both processors, then they also don't
show the advantage of boards like the S2885 with DIMM slots for
each processor.

That's the trouble
with these sites, they get all the cool hardware but do games tests on
these boards. Who cares? Slap a high-end RAID SCSI or SATA card in and
see if the board locks-up under intense IO. Look for problems, not just
make pretty graphs. I know that the system I'm using now, a AMD MPX based
Gigabyte GA-7DPXDW+ certain has it's shares of foibles that none of the
hardware reviews I've seen ever found.

Which is why I like to browse the Asus, Tyan, etc newsgroups
before trying a motherboard I've never used before.

I've certainly noticed how few problems with the Tyan Opty
dualies are showing up in the newsgroups as compared to problems
with non-Tyan boards.

Personally, the only real problem I've had with the S2885 is
that the 4-port SIL3114 SATA controller allows RAID 0+1, but
it seems to be impossible to boot from a RAID on that controller.
Ended up using an 8-port Adaptec PCI-X card in that system.