Desktop C2D motherboards supporting more than 8GB?

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Topic title says it all. With the introduction of C2Q I can see where 16GB or 32GB would be extremely useful especially for running multiple VMs.
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
D'oh. I found renethx's excellent Core 2 thread which indicates that the only motherboards supporting >8GB RAM are based on nVidia's nForce4, nForce 570, nForce 590, or nForce 680i SLI (up to 32GB) chipsets.

It looks like >8GB support will become widespread in 2007.
 

Roguestar

Diamond Member
Aug 29, 2006
6,045
0
0
Phwoah. Assuming that 99% of motherboards don't have more than 4 DIMM slots, that means to get >8Gb you need sticks of RAM of greater capacity than 2Gb each. And 4x 4Gb sticks of DDR2 is going to bring down the house in terms of cost!
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: Roguestar
And 4x 4Gb sticks of DDR2 is going to bring down the house in terms of cost!
True today. 2 x 1GB DDR2 is pretty reasonable now; 4GB sticks will be reasonable in 18 months.
 

BladeVenom

Lifer
Jun 2, 2005
13,365
16
0
I doubt it. 2GB sticks might be reasonable in 18 months, but I would guess 4GB will cost about what 2GB does now or more.
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
More than two slots per channel would mandate Registered or Fully Buffered DIMMs, and that in turn forces you away from desktop class to server/workstation chipsets.
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: Peter
More than two slots per channel would mandate Registered or Fully Buffered DIMMs, and that in turn forces you away from desktop class to server/workstation chipsets.
Will we see non-ECC, non-registered, non-buffered (i.e. desktop) DDR2 memory in the 4GB and 8GB per stick range?

Two questions really: first, is it technically possible? Second, will the market demand such products?

The latter question is not simply theoretical. I know that memory makers are moving to DDR3. Will there be enough demand for 4GB/8GB DDR2 memory to make it worth their while?
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
Two-gigabit _chips_ are planned for DDR2, so 4-GiB unbuffered DIMMs (16 chips w/o ECC, 18 w/) would be feasible.

Since it's the chip addressing that limits the DIMM size for a given chipset, you'll have to look at whether today's chipsets actually support chips that large.
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: Peter
Since it's the chip addressing that limits the DIMM size for a given chipset, you'll have to look at whether today's chipsets actually support chips that large.
I'm assuming that the chipsets I mentioned in the second post provide such support. Is that a fair assumption?

On the other hand, not all motherboard makers seem to try to take advantage of that opportunity.
 

oldgeezzer

Junior Member
Dec 5, 2006
3
0
0
2^^32 allows only a little over 4 gig byte addressing. A lot of the "how much memory is enough" problem depends on the OS being used
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: oldgeezzer
2^^32 allows only a little over 4 gig byte addressing. A lot of the "how much memory is enough" problem depends on the OS being used
Good point. I'm thinking 64-bit Windows Vista Business or Ultimate, both of which support up to 128GB RAM.
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
oldgeezzer, it's actually /exactly/ 4GiB address space, and a lot less RAM (because there's other stuff that needs to appear in that space), typically 3.5 to 2.5 GiB actually system-visible RAM - mostly depending on how much space your graphics card and other I/O use.
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
The reason I'm interested in machines with >8GB is that I plan to run multiple virtual machines, e.g. Virtual PC 2004 or VMware's Virtual Server. These products cry out for real RAM. In addition, my video editing software loves lots of RAM too.

When you add the ability to run quad cores in a desktop machine then the opportunity to put >8GB to work seems obvious.
 

severtki

Member
Apr 9, 2003
177
1
71
Cheaper solution: Build several separate machines and use KVM switch or VNC server to control them. You could build yourself 3 or 4 cheap machines, each with 2 GB RAM for lots cheaper than outfitting a single quad-core machine with 8GB+ RAM.

Just a thought... probably not what you're looking for, but...
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
NVidia's public specs are all buzz and little facts, so I can't really tell whether these chipsets do or do not support 2-gigabit RAM chips. Not supporting 1-gigabit types would be rather pathetic, so I'd firmly expect that to be OK.
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: severtki
Cheaper solution: Build several separate machines and use KVM switch or VNC server to control them. You could build yourself 3 or 4 cheap machines, each with 2 GB RAM for lots cheaper than outfitting a single quad-core machine with 8GB+ RAM.

Just a thought... probably not what you're looking for, but...
LOL. That's exactly what I've got now: 3 x Athlon XP 2x00+ based machines along with a couple older workstations. I plan to assemble 3 (probably C2D upgradeable to C2Q) based machines in 2007. But a max of 8GB ain't going to cut it for any of those machines especially running a 64-bit OS.
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
Keep in mind that C2D's total address space is only 36-bit (64 GiB), Opterons currently use 40 (one Terabyte).

If you want to build something today that lets you upgrade next year or even later, then you should REALLY opt for an Opteron box - these have the RAM controllers on the CPU, which support large DIMMs, and also bring another pair of RAM controllers with every CPU socket ... and have a longevity plan that lets you put future CPUs into today's sockets for quite a while longer than Intel will let you.

So a dual-socket Opteron could get you started with one dual-core CPU and two DIMMs, and take you up to eight cores, four RAM channels and up to eight or sixteen DIMMs ... registered ECC DDR2 though.

Like, if you're worried about max RAM size, you might find this here quite nice ;)

http://www.tyan.com/products/html/thunderh2000m.html

 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: Peter
Keep in mind that C2D's total address space is only 36-bit (64 GiB), Opterons currently use 40 (one Terabyte).
I wasn't aware of this; thanks for pointing it out. 64GB would satisfy my needs. Now if vendors would only implement even close to that.
If you want to build something today that lets you upgrade next year or even later, then you should REALLY opt for an Opteron box - these have the RAM controllers on the CPU, which support large DIMMs, and also bring another pair of RAM controllers with every CPU socket ... and have a longevity plan that lets you put future CPUs into today's sockets for quite a while longer than Intel will let you.
Frankly I'm not convinced that AMD, even with K8L, will be able to *convincingly* overtake C2* in 2007. I think AMD is going to be competitive and will be able to preserve their critical gains in the server space starting in the 2007H2. That doesn't help me; I need a solution starting in the first quarter of 2007 and C2* seems more than sufficient for my needs. Current overclocking convinces me that there's plenty of clock "headroom" in the C2* architecture.
So a dual-socket Opteron could get you started with one dual-core CPU and two DIMMs, and take you up to eight cores, four RAM channels and up to eight or sixteen DIMMs ... registered ECC DDR2 though.
Workstation class boards, dual sockets, ECC memory, ahhh... that's the rub. Too rich for my budget.

With the advent of dual and quad core processors, multiple physical processors become extremely uninteresting to me. I can get more than adequate performance with 4 cores and 16GB or 32GB of RAM
Like, if you're worried about max RAM size, you might find this here quite nice ;)

http://www.tyan.com/products/html/thunderh2000m.html
:p Oh the fun I could have with one of these puppies fully tricked out in a snow bound cabin with Virginia Madsen by my side. Come to think of it, who needs the hardware - software rulez.
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
Why still multiple CPUs? With Intel's northbridge centric approach, there's little point. With AMD, you have the choice - many cores working on one pair of RAM controllers, or many nodes, each with its local pair of RAM controllers. The latter is performing considerably better when you have tasks that work on separate datasets - as you would when you're running virtual machines - and of course lets you have more DIMMs overall. Also, you don't /have/ to populate both processors in a dual socket board. Buy one now, upgrade later - add a second dualcore CPU and another pair of DIMMs, or maybe even a quad CPU alongside the dual CPU you originally had, why not.

There might be plenty of headroom in C2, but just look at the plans (and the historical backlog) to figure how far an Intel board bought /today/ is going to take you. Hint: Not too far.

AMD in turn have clearly indicated that at least the next two generations of Opterons will fit the current socket.

As a side note, max RAM size with AMD's current CPUs appears to be 4 GiB per chip (should they ever be made) - that translates to 8 GiB per unbuffered DIMM, sixteen per registered DIMM. Four unbuffered DIMMs take you to 32 GiB on an AM2 desktop board, eight registered DIMMs on a workstation board take you up to 128 GiB ... per CPU node, 512 GiB total for four of them.

That's why AMD is selling their Opterons like hotcakes, despite C2D being a /really/ good processor. It's the scalability and the longevity of stuff bought today.
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
Another side note: AMD's "4x4" effort for performance gaming will bring more mainstream oriented dual-socket Opteron-ish boards to the market Real Soon Now. Have a look ...
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: Peter
Why still multiple CPUs? With Intel's northbridge centric approach, there's little point. With AMD, you have the choice - many cores working on one pair of RAM controllers, or many nodes, each with its local pair of RAM controllers. The latter is performing considerably better when you have tasks that work on separate datasets - as you would when you're running virtual machines - and of course lets you have more DIMMs overall. Also, you don't /have/ to populate both processors in a dual socket board. Buy one now, upgrade later - add a second dualcore CPU and another pair of DIMMs, or maybe even a quad CPU alongside the dual CPU you originally had, why not.
I've ruled out dual processor boxes (i.e. boxes with two CPU sockets) due to motherboard and memory costs. In addition, Anand's recent review of C2D vs. C2X vs. QFX doesn't show QFX in a good light for general computing (i.e. non-server) workloads.
There might be plenty of headroom in C2, but just look at the plans (and the historical backlog) to figure how far an Intel board bought /today/ is going to take you. Hint: Not too far.
This is interesting speculation. Intel ruled for so long we've taken their arrogance for granted. Having spent 2002 through mid-2006 getting the stuffing kicked out of them seems to have had some positive consequences. C2* is a real *asskicking product line (certainly in the mobile and desktop markets) with real growth (i.e. performance) potential. This is not your father's Pentium4 with the Rambus memory mumbo-jumbo.

AMD in turn have clearly indicated that at least the next two generations of Opterons will fit the current socket.
It seems to me that performance of AMD's next generation is going to *try* to beat Intel's currrently shipping products. We know Intel's current architecture can ramp up to 4GHz range.; it's not clear if AMD's next generation will be able to do so. The current QFX offering for desktop-type tasks gets blown away by C2X. As for AMD's architecture after K8L, who knows...

As a side note, max RAM size with AMD's current CPUs appears to be 4 GiB per chip (should they ever be made) - that translates to 8 GiB per unbuffered DIMM, sixteen per registered DIMM. Four unbuffered DIMMs take you to 32 GiB on an AM2 desktop board, eight registered DIMMs on a workstation board take you up to 128 GiB ... per CPU node, 512 GiB total for four of them.

That's why AMD is selling their Opterons like hotcakes, despite C2D being a /really/ good processor. It's the scalability and the longevity of stuff bought today.
On server platforms, I agree with you 100%. I'm not interested in building server-class boxes.

Remember that my purchase has a 3 year life: 2007 through 2009. Based on what I see right now in the desktop class, it's going to take AMD that long to prove they can surpass C2*. QFX is looking pretty lame right now for desktop-type workloads and K8L is a big question mark and will be generally available 2007Q3 at the earliest. I can't wait that long.

My original plan was to assemble 3 identical platforms in the first half of 2007, starting with motherboards that are C2D/C2X/C2Q compatible and populating them with E6400 level processors. As the C2* product line unfolds I have the option to upgrade to C2X or C2Q. That's a sweet upgrade path and it exists today. There's nothing in AMD's current offerings that make me want to revise that. K8L, well, that's 6-9 months away for all practical purposes.
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
Let me make one thing clear: You ARE asking for a server-class box, and you're not going to get what you want from a desktop platform.

Using a C2D on a desktop board might get you the nicer (I said that!) processor, but on a platform that is weaker than you want TODAY.

That's why I recommended doing the reverse: Use more of today's money for a mainboard architecture that gives you the platform headroom you need, and which coincidentally will let you use the next two future (!) generations of processors and more RAM than you'll ever want.

Ask yourself: In 2008, is 2006's C2Q on a desktop board, limited to 16GB RAM (at most!) still going to cut it? Besides, I wouldn't expect Intel to keep today's top-of-the-line offerings available until then. FSB change, chipset compatibility "issues", and off you go buy a new mainboard to your new CPU. That's how it's always been since the 486 days, and that's how it's going to continue as long as there is a front side bus/northbridge architecture.

Reg ECC RAM isn't /that/ more expensive than /quality/ unbuffered, btw. Intel will make you buy Fully Buffered DIMMs next year, and these /are/ expensive.
 

rhatsaruck

Senior member
Oct 20, 2005
263
0
0
Originally posted by: Peter
Let me make one thing clear: You ARE asking for a server-class box, and you're not going to get what you want from a desktop platform.
I agree that my requirement (i.e. motherboards that support >8GB) is traditionally considered to fall in the server-class hardware space. However with the widespread availability of 64-bit processors, multi-core processors, and 64-bit operating systems (Vista and any flavor of Linux) the traditional boundaries between desktop, workstation, and server class spaces are changing. I am trying to take advantage of that fact and yet stay within the desktop space. There's a premium to be paid in the workstation and server markets. I specifically don't want to pay that premium. And although I haven't mentioned my other requirements I can share with you that they fall solidly within the desktop space.

Using a C2D on a desktop board might get you the nicer (I said that!) processor, but on a platform that is weaker than you want TODAY.

That's why I recommended doing the reverse: Use more of today's money for a mainboard architecture that gives you the platform headroom you need, and which coincidentally will let you use the next two future (!) generations of processors and more RAM than you'll ever want.
AMDs K8L/"Stars" class processors for the desktop are not due until 2007Q3. It's clear from this article that current motherboards will *not* be able to take full advantage of at least some of the capabilities of those processors. Why in the world would I invest in a platform (AM2) that requires me to wait 6-9 months (i.e. until 2007Q3) to discover if a new processor architecture can beat what Intel's been shipping since 2006Q3?

Ask yourself: In 2008, is 2006's C2Q on a desktop board, limited to 16GB RAM (at most!) still going to cut it? Besides, I wouldn't expect Intel to keep today's top-of-the-line offerings available until then. FSB change, chipset compatibility "issues", and off you go buy a new mainboard to your new CPU. That's how it's always been since the 486 days, and that's how it's going to continue as long as there is a front side bus/northbridge architecture.
I'd be perfectly happy with a motherboard that takes 16GB RAM; 32GB would be better but *any* improvement over 8GB is better than nothing. I don't need the fastest/bestest/b*itchenest systems on the block. I'm looking for a platform that lets me move up from C2D-level of price/performance (my starting point in 2007Q1) to C2Q-level of performance sometime in late 2008 when, I expect, C2Q absolute processor prices will drop. C2*'s price/performance is looking pretty good today. We know the C2* performance can be pushed to the 4GHz level - people are overclocking today's motherboards to get that level. If folks at the various enthusiast sites can make it happen for a few hundred dollars I'm willing to bet that Intel can figure it out too. ;)

Reg ECC RAM isn't /that/ more expensive than /quality/ unbuffered, btw. Intel will make you buy Fully Buffered DIMMs next year, and these /are/ expensive.
I think server-class RAM is traditionally at least 25% more expensive than desktop RAM. Why should I spend 25% more and not get any benefit based on my needs.

Peter, I really appreciate your comments. They are helping me clearly identify my needs and think through my decision-making process. Thank you.
 

Peter

Elite Member
Oct 15, 1999
9,640
1
0
16 gigs on desktop C2D means four-gigabyte unbuffered DIMMs. These, should they ever be made and should chipsets for Intel processors actually support them, will be insanely expensive, much more expensive than 2- or 4-GiB registered DIMMs will ever be.

That is because reg DIMMs can be made from twice as many chips, eliminating the need for the expensive 2-gigabit chips. Just look at the price gap between 1- and 2-gig unbuffered DIMMs right now to get the idea. Prices from a big retailer here in Germany: Unbuffered (sixteen expensive 1-gigabit chips) 400 euros, registered ECC (32+4 pieces of mainstream 512-megabit RAM chips) is 260. Same speedgrade, DDR2-667.

So, cost advantage of desktop technology? Hardly. On RAM alone, workstation stuff is 560 euros CHEAPER (about 700 US$) - and that's calculating eight gigs, not the sixteen you want. $700 buys you one sweet mother of a workstation board AND an extra dualcore CPU - right now.

Besides, with these amounts of RAM, you /want/ ECC ... the more RAM you have, the more practically relevant does the effect of "soft errors" become.

Past experience also suggests that you won't be able to buy a processor upgrade for a late-2006 Intel board in late 2008 at all. AMD are also set to get some improvement into their Socket F and make it F+ or something, but they also have officially announced that it'll be compatible both ways. Sure, you won't be able to use the improvements (like, apparently, more HT links and higher HT speeds), but in return you don't have to buy a new board, new RAM, new power supply and whatever else might be in store, just to get a CPU upgrade in. Like it or not, future proof and Intel in the same sentence just does not compute - particularly not in the desktop segment.

Thus, if you go with C2D and desktop technology, you'll be stuck with 8 gigs of RAM, and you'll be stuck there or starting a new system build much sooner than you'd like to ... without being able to reuse your expensive RAM.

Of course it's your money and you're free to do whatever you please. You have been warned, and - I'm glad you said that - you have been informed too ;) Sum up the actual cost of the entire thing, ring the total, and you'll see what I mean ... even without the future outlook.