Commoditization of 4-way

[Tony Hill]

The price of a 4-way system is not much over $5000 *now*, isn't it?
$1700 for the Tyan board, $2800 for four Opteron 844's, $250 for a chassis,
and, well, you get one SATA disc and only one of the Opterons gets any memory
with the change from $5000.

Even with such an incredibly stripped-down system you're still looking
at $5000, but for all practical 4P Opteron systems you're looking at
about $10,000 as a base price. Also, Opterons are very reasonably
priced as far as 4P systems go. XeonMPs will tend to cost you more,
and the price only goes up from there when you start looking at
Itaniums, IBM Power chips, Sun SPARC chips, etc.

FWIW the absolute cheapest that HP will sell you a 4P Opteron server
for is $11,297. That gets you 4 Opteron 842 (1.6GHz) chips and 2GB of
memory, For comparison, the absolute cheapest 4P Intel XeonMP
server that HP sells will set you back $11,265, but there you only get
4 2.0GHz XeonMP chips and 1GB of memory. With Itanium things start
getting REAL pricey. A bare-bones 4P Itanium2 systems from HP will
set you back $31,555, and that only gets you 4 x 1.3GHz/3MB L3 chips
and 1GB of memory.

I wonder if AMD will drop the Opteron 840 price enormously at some stage;
4 x 1400MHz with separate memory to each processor isn't bad for, say, a
shell-account machine.

Highly unlikely. They've already got their 840, 842 and 844 all
priced exactly the same, $698. VERY reasonable price as compared to
the competition, but very unlikely to drop any further. Normally when
you start seeing price parity with higher speed chips like that it
means that the slower models are in the process of being discontinued.

 

[Tony Hill]

$1545 at Lynn Computer, up from $1495 a couple of weeks ago.
They are taking orders only - no stock yet.
http://www.lynncomp.com

Ahh the infamous Lynn Computers. They ALWAYS take orders for thing,
regardless of whether or not they plan on getting them in stock! That
store is constantly listing products that won't ship for months.

 

[Tony Hill]

When I can get a 4-way Opty MB for about $500, and the
CPUs for about $350 apiece, I'll consider the price almost
in my range.

Hopefully that time might be coming soon, that's what started this
whole thread. The trick is that next year AMD is expecting to ship
dual-core Opterons that are pin-compatible and
infrastructure-compatible with existing single-core Opterons. If all
goes according to plan they will be drop-in replacements for current
Opterons, so all you would need for a 4-core machine is a "2
processor" motherboard (whether or not that makes a true "4 processor"
system or just a "2 processor with dual cores" machine depends on your
point of view).

Those 2P motherboards start at only ~$200, though good ones will cost
you more like $500. No word yet on the price of the chips, but they
might sell for $350 a piece, and most likely will sell for less than
$700 a pop (2 cores for $700 will keep in your price range of
single-core for $350).

 

[Tony Hill]

Yawn - the 21364 had that how many years ago?

Considering that the 21364 only just started shipping a few months
before the Opteron... not much (Jan. 2003 for the Alpha EV7, April
2003 for the Opteron). Also there is a grand total of *ONE* system
being sold with the Alpha EV7 processor, the HP GS1280, and it'll set
you back a cool million dollars or so.

In short, not really a worthwhile comparison. For all intents and
purposes, the 21364 is a non-existent product.

 

[Greg Lindahl]

In short, not really a worthwhile comparison. For all intents and
purposes, the 21364 is a non-existent product.

Can we please stop wasting everyone's time by telling other posters
that their example is irrelevant because it doesn't ship in enough
volume, etc? It's pointless.

-- greg

 

[nobody]

On 18 May 2004 23:59:57 +0100 (BST), Thomas Womack

....snip...

I'm not sure I'd call a system with two dual-core CPUs a quad system,
though I'm not quite sure where that prejudice comes from; I suppose
that part of the issue of a quad system is the enormous motherboard
required physically to fit four sockets, four cooling systems, four
sets of memory ... on memory-intensive tasks I think I'd rather have
more memory subsystems than more cores, dual-core Opterons will be no
less memory-starved than 800MHz FSB Noconas.

Tom

With each dual-core chip containing the same dual channel memory
controller as the current crop of socket 940 chips, each core would
have the same memory bandwidth as current socket 754 Athlon64. While
not quite as impressive as A64 FX/Opteron, A64 xx00+ is still a
formidable CPU and doesn't seem to really badly suffer from
insufficient memory bandwidth. If I could get a quad A64 on the cheap
(comparably priced to the dual Opteron 242/MSI Master2-far I am
building now), I'd go for it without much thinking. Unfortunately
it's not possible technically (the number of HT links enabled on each
CPU etc.) But 2 dual core Opterons would closely resemble that
hypothetical quad A64, just better because it would have 2
heatsinks/fans less.
As for having more memory subsystems than cores, here is an article
comparing lowly (among dual boards) MSI K8T Master2-far to higher end
Tyan K8W Tiger and yet even higher Tyan K8W Thunder.
http://www.neoseeker.com/resourcelink.html?rlid=68739
Contrary to expectations, the more expensive Thunder with both CPUs
connected to own memory is not any better than its humble competitors
with one CPU accessing the memory through the other (give or take a
fraction of a percentage point).

 

[Rob Stow]

Ahh the infamous Lynn Computers. They ALWAYS take orders for thing,
regardless of whether or not they plan on getting them in stock! That
store is constantly listing products that won't ship for months.

If *any* vendor doesn't have something in stock I don't order
from them. At best I'll ask them to call or e-mail me when
they do actually have something to sell. And with any vendor,
if the product isn't at my door within 24 hours of the promised
delivery date I call in to cancel the order.

As far as Lynn goes, I've ordered three Tyan S2885 motherboards
and 6 Opty processors to put on those - two orders altogther -
and both orders arrived 3 days later. Not that bad for cross
border shipping (US to Canada). I suspect it takes 24 hours to
get to Regina and then another 48 hours to cover the last 70 km
from Regina to Moose Jaw.

 

[Greg Lindahl]

Contrary to expectations, the more expensive Thunder with both CPUs
connected to own memory is not any better than its humble competitors
with one CPU accessing the memory through the other (give or take a
fraction of a percentage point).

Given that you didn't give any details at all, the most likely
explanation is pilot error. I know that STREAM on Linux 2.6 with the
bios interleave set off and a user-level utility to confine the
processes appropriately is much improved. And my testing with
various HPC codes also showed significant improvement.

-- greg

 

[myren, lord]

i agree that the 2 core opterons will likely open the floodgates on
4-way heaven and begin the price falls.

 

[Yousuf Khan]

Hopefully that time might be coming soon, that's what started this
whole thread. The trick is that next year AMD is expecting to ship
dual-core Opterons that are pin-compatible and
infrastructure-compatible with existing single-core Opterons. If all
goes according to plan they will be drop-in replacements for current
Opterons, so all you would need for a 4-core machine is a "2
processor" motherboard (whether or not that makes a true "4 processor"
system or just a "2 processor with dual cores" machine depends on your
point of view).

Those 2P motherboards start at only ~$200, though good ones will cost
you more like $500. No word yet on the price of the chips, but they
might sell for $350 a piece, and most likely will sell for less than
$700 a pop (2 cores for $700 will keep in your price range of
single-core for $350).

That's another thing, dual-core 2-way processors. However, I think these
server makers are actually looking to getting down the price of a true 4-way
processor system down to the $5000 range.

Yousuf Khan

 

[nobody]

Given that you didn't give any details at all, the most likely

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Have you not noticed the link? Just in case I am providing it here
again: http://www.neoseeker.com/resourcelink.html?rlid=68739
Please note that I bear no responsibility for either the content of
that article or the methodology the authors applied to their
benchmarks. I just found the comparison interesting and worth noting
in the groops.

 

[del cecchi]

snip.

As far as I can see the main difference in practice right now is that
there is a PCI-E connector and there isn't one for HT. And the error
handling issues Del noted (hopefully be fixed with HT 2.x)
Ok and HT hotplug would be nice too.

-Andi

Can't have much wire and a connector in a 56 ps skew+jitter allocation
(at 1.6 Gb/s), with no provision for aligning clock and data. :-(
The last draft of V2.0 goes to 2.8 Gb/sec still with no alignment. I'll
have to look at the "networking extensions", but the 8131 doesn't have
any recovery. Maybe there is a new version coming.

del

 

[KR Williams]

Can we please stop wasting everyone's time by telling other posters
that their example is irrelevant because it doesn't ship in enough
volume, etc? It's pointless.

Is it any more pointless than to say "XYZ" did it first, but
failed in the marketplace? ...other than in the .folklore group?

 

[Stephen Fuld]

i thought intel was one of the infiniband people too.

Initially they were. There was a "falling out" between the server people,
who wanted the advantages of IB and the desktop people who, while they were
initially on board, saw it getting too complex and too far delayed for their
tastes and decided to go with their own solution. That was when Intel
decommitted doing IB in the chipset.

infiniband and
pci-e arent mutually exlcusive technologies by any means.

I'm not sure what you mean here. If you mean you can have some sort of PCI
adaptor that talks IB on the other side, then yes, but I wouldn't consider
that to be "real Infiniband". A big part of the advantage of IB is the
elimination of the PCI, memory mapped I/O scheme, etc. and having the
interface directly in the chipset. If you mean you could have a chipset that
supports both, then yes, I suppose you could, but I don't think anyone will,
as they are pretty much competitors for that function.

parents said amd wasnt a fan of pci-e. are they designing any competition?

I don't know anything about what AMD is doing.

 

[Stephen Fuld]

Seen from the software (not firmware) side HT is basically completely PCI
compatible. I don't know of any visible differences. The north
bridge on a Opteron system is implemented in the CPU and e.g. all the
PCI config accesses originate from the north bridge. But the request
has to travel over an HT link before it can actually talk to an real
PCI bridge. This works completely transparent.

Then ISTM that HT and PCI are at a different level. i.e. if you have HT to
PCI, then PCI and HT aren't direct competitors. This jibes with my,
admittedly modest, understanding of HT as primarily an interchip (CPU-CPU
and CPU-"Northbridge") interconnect and PCI as primarily a way to connect
external peripheral interfaces such as Ethernet and Fibre Channel/SCSI.
Thus there isn't a decision between HT and PCI, as each have their potential
(though different) place in a system.

 

[Greg Lindahl]

Contrary to expectations, the more expensive Thunder with both CPUs

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Have you not noticed the link?

Yes, I read the link. It didn't say anything about the BIOS settings
or whether the OS was told/has the capability of restricting processes
and their memory to particular cpus. As you can see from my
explanation of how I saw better performance, these details all matter.

-- greg

 

[Andi Kleen]

Yes, I read the link. It didn't say anything about the BIOS settings
or whether the OS was told/has the capability of restricting processes
and their memory to particular cpus. As you can see from my
explanation of how I saw better performance, these details all matter.

An NUMA aware OS with the right BIOS settings (no node interleave)
should be able to run multithreaded STREAM even without explicit tuning,
giving near perfect scaling on Opteron systems with local memory
on each CPU. Linux does usually.

-Andi

 

[Tony Hill]

Can we please stop wasting everyone's time by telling other posters
that their example is irrelevant because it doesn't ship in enough
volume, etc? It's pointless.

It's no more pointless than pointing to the 21364 as an example of
much of anything beyond a proof of concept. One pointless message
answered by another? Perhaps.. welcome to Usenet.

 

[del cecchi]

Initially they were. There was a "falling out" between the server people,
who wanted the advantages of IB and the desktop people who, while they were
initially on board, saw it getting too complex and too far delayed for their
tastes and decided to go with their own solution. That was when Intel
decommitted doing IB in the chipset.

Some Intel folks are still involved in the trade association. But not
to the extent they were before.

del

 

[Peter Boyle]

It's no more pointless than pointing to the 21364 as an example of
much of anything beyond a proof of concept. One pointless message
answered by another? Perhaps.. welcome to Usenet.

Oh, for goodness sake.

And it was Ford who invented the motor car, since noone else
mass produced, right?

Peter