AMD 4x4-platform based on Socket 1207 / AMD roadmap ?

Printable View

Show 100 post(s) from this thread on one page

09-24-2006, 12:02 PM
nn_step

Quote:

Originally Posted by awdrifter

Then what's the advantage of registered ram? Why do servers use them? Thanks.

Registered memory is a specific form of memory designed to support a larger number of chips on a given memory module. To allow for all the memory chips on the module to properly function, it is necessary for the module to add a delay to all command requests to the memory. The delay is actually generated by an additional clock generator that resides on the module instead of relying on the motherboard. This will decrease the overall performance of a given memory module. The memory type is typically used in large server environments where large amounts of RAM are required. In these environments, the benefits of the increase in RAM outweigh the performance impacts caused by memory swapping.

Quote:

Originally Posted by Teroedni

Yea but what about overclocking?

:rofl: :lol: :rofl: :imo:
09-24-2006, 12:20 PM
[XC] Teroedni

Okey
It seems it is actually possible to overclock 1207 server boards if their bios supports it.
So a 4x4 certified board should be able to overclock well then:D

I guess Asus can just rebrand a server board like the Asus KFN32-D SLI and just add some bios options and voila 4x4;)
09-24-2006, 12:37 PM
cky2k6

Quote:

Originally Posted by nn_step

Registered memory is a specific form of memory designed to support a larger number of chips on a given memory module. To allow for all the memory chips on the module to properly function, it is necessary for the module to add a delay to all command requests to the memory. The delay is actually generated by an additional clock generator that resides on the module instead of relying on the motherboard. This will decrease the overall performance of a given memory module. The memory type is typically used in large server environments where large amounts of RAM are required. In these environments, the benefits of the increase in RAM outweigh the performance impacts caused by memory swapping.

actually, registrered ram doesnt have too much of a penalty on latency, thats ecc, which basically just ads one more clock cycle, so cas 2 ram is cas 3 ecc. registered non-ecc ram will actually function at the same latency as it was non-buffered, its just that nobody has really cared enough to make non-ecc ram registered, as 32 bit os's cant use registered ram anyways, as you can have up to 16gb of ram as u-buffered. registered ram will be a little slower, but not a whole clock cycle slower like ecc.
09-24-2006, 02:09 PM
arisythila

I believe what he was meaning. Is there is no decent Registered memory out.. Fastest we've seen in 1 gb x 2 is DDR2-667. Thats crap in my opinion. I want to see some nice DDR2-1000 that is overclockable to DDR2-1200, like we can with regular DDR2 memory. This is what I want.. I just hope if they take the route of going to Scoket 1207.. They make it so we can use normal DDR2 memory.

~Mike
09-24-2006, 03:47 PM
nn_step

Quote:

Originally Posted by cky2k6

actually, registrered ram doesnt have too much of a penalty on latency, thats ecc, which basically just ads one more clock cycle, so cas 2 ram is cas 3 ecc. registered non-ecc ram will actually function at the same latency as it was non-buffered, its just that nobody has really cared enough to make non-ecc ram registered, as 32 bit os's cant use registered ram anyways, as you can have up to 16gb of ram as u-buffered. registered ram will be a little slower, but not a whole clock cycle slower like ecc.

What makes you think you can have non-ECC registrered ram?
09-24-2006, 04:13 PM
***Deimos***

Quote:

Originally Posted by nn_step

What makes you think you can have non-ECC registrered ram?

good point

registered/buffered memory primarily serves the purpose of reducing the huge capacitative load on the memory controller allowing more memory to be installed. (for effect of capacitative load on memory controller, compare overclock/timing of single sided DIMM vs double sided, and 1 DIMM/channel vs 2 DIMMs (ie 4 total).

http://en.wikipedia.org/wiki/Registered_memory
09-24-2006, 04:58 PM
The Ghost

Quote:

Originally Posted by Nedjo

Samsung has Registered DDR2-800:

http://www.samsung.com/products/semi...93T2950EZA.htm

nice find , the first that i have seen :)
09-24-2006, 05:20 PM
cky2k6

Quote:

Originally Posted by nn_step

What makes you think you can have non-ECC registrered ram?

because they two techs are not related... ecc is for error correction, registered ram is for capacity. theyre only found together since theyre both critical on servers, and useless so far for enthusiasts... there has been non-ecc registered ram in the past, not common, but it exists...
09-24-2006, 05:20 PM
oldblue

This is what Rahul Sood, founder and CTO of VoodooPC, had to say:

Quote:

With the help of Nvidia, and others, AMD is creating an entirely new motherboard concept for 4x4. It's not your standard-issue workstation motherboard; rather, it's an enthusiast-class multiprocessor motherboard with support for unbuffered non-ECC enthusiast-class memory. AMD also promises to release three tiers of 4x4 in the box where processor kits will start well under $1,000 and go up from there.

http://voodoopc.blogspot.com/2006/09...tegory-of.html
09-24-2006, 06:34 PM
Fred_Pohl

Just because 4x4 will be using socket 1207 doesn't mean that 4x4 will require registered ram.
09-24-2006, 08:34 PM
***Deimos***

I guess people are being haunted by memories of the initial FX series (ie registered RAM, socket 940)

Personally, I can't wait and see what AMD delivers with "socket" 1207. Reserved pins for more HT links? Reserved pins for other memory types? Very curious to see how many DIMMs new Opterons will be able to push. I bet that memory controller is built like a tank.
09-24-2006, 09:03 PM
nn_step

Not to mention the native DDR3 support :D
09-25-2006, 02:12 AM
Nedjo

HKEPC view on the subject (they are claming that the source is mobo manufacturers...):

http://www.hkepc.com/bbs/itnews.php?...me=0&endtime=0
09-25-2006, 04:20 AM
el rolio

Quote:

Originally Posted by ***Deimos***

I guess people are being haunted by memories of the initial FX series (ie registered RAM, socket 940)

Personally, I can't wait and see what AMD delivers with "socket" 1207. Reserved pins for more HT links? Reserved pins for other memory types? Very curious to see how many DIMMs new Opterons will be able to push. I bet that memory controller is built like a tank.

yea i was talkin about this a few pages back. i mean it could be a good idea but what we really need is MORE INFO FROM AMD
09-27-2006, 07:19 AM
uOpt

Quote:

Originally Posted by awdrifter

Sorry for the noob question, but what's so bad about registered rams? Are they not as fast?

There is only a minimal overhead for the registered and at this time almost nonexistent for the ECC. You can easily test the latter by turning on and off ECC in a board with unregistered RAM. I might have some numbers somewhere.

The problem is overclocking RAM. Overclocking RAM is a product made from selected chips, or at least selected modules.

The enthusiast RAM vendors would hate to, and probably will not, offer an overclocking registered product. That didn't even happen when the first Athlon FXes were the craze and required registered DDR. OCZ and Mushkin made some low-latency but that was it.

The sad thing is that it is all rubbish anyway. RAM speed only has a minimal impact on real-world performance and that "run 1:1 for best overclock" rule seems to be a myth, too (at least my max dual mprime stabe clocks don't drop when I take down the RAM).

But the overclocking crowd doesn't care. They want fast RAM anyway.

Quote:

Originally Posted by turtle

If 4x4 is on 1207 and still requires buffered (registered) ram, what exactly is the point of trying to sell it over a dual-socket server board with 22xx Opterons which would be the exact same thing?

The only point would be the unlocked multiplier and the faint hope that board vendors offer overclocking boards for dual 1207.

Quote:

There's got to be something cooking on that front. Even if not, at least 1207 doesn't require ECC...I think a lot of people confuse the two.

All current processors support but no not require ECC, including AMD64, Core2, PowerPC and Itanium.

It would in fact be possible for the enthusiast RAM vendors to make registered non-ECC to cut down cost and maximize overclock potential for socket 1207 if they think there's a market of people buying it (there isn't).

Quote:

Originally Posted by awdrifter

Then what's the advantage of registered ram? Why do servers use them? Thanks.

You reduce the load on the control lines so that the same processor can drive more chips. As a result the same mircoprocess can drive twice as many modules and each module can have twice as many chips, resulting in a 4x increase of memory capacity. It also reduces the change of RAM error from random electrical sources since the signal is now stronger.

Keep in mind that in registered RAM as in for socket603/ 604, 940 and 1207 only the control lanes are buffered, the data lanes are not. FB-Dimms buffer the data, too, but FB also changes the whole setup to a cascading one.

Quote:

Originally Posted by Fred_Pohl

Just because 4x4 will be using socket 1207 doesn't mean that 4x4 will require registered ram.

Read the disription of the pin layout. The socket 1207 platform is for registered RAM.

The memory controller is on the CPU for AMD64s, so there no switching between registered and unregistered with the same CPU and/or socket.

In the Intel world that would work, because the CPU doesn't have the memory controller. The same Intel CPU can theoretically use registered and unregistered RAM in different boads. But in practice they are locked, too, because Intel usually only relases one chipset, and then the chipset fixes what kind of RAM you need.

Or in other words, you could run a Woodcrest with DDR2 unbuffered or non-FB registered, as in the CPU can, but you cannot buy a chipset to do so.
09-28-2006, 05:37 AM
Poodle

What's all this chatter about ECC/reg memory? :stick:

Hey guys. It would be pointless of introducing 4X4 with anything other than normal unreg/Non ECC memory modules because if they wouldn't it would be just a normal Socket F system.

Last year I believe that Abit showed of an Opteron socket 940 (e.g. not 939) board that didn't make it to the shops but seemed to work, that could handle unreg/non ECC. So why wouldn't all the big guys over at Amd not be able to get this to work?

Amd wouldn't sell any 4x4 if they wouldn't. We know it and they must surely know it! If i'm wrong the 4x4 was just a media stunt to get people not to jump on Kentsfield.
09-28-2006, 05:54 AM
lapdog

These FX chips must be different than Socket F. If they are the same they will sell like hotcakes to potential Opteron buyers as a single 2.6GHz Opteron 285 is over $900. I am sure the memory controller is the only difference.
09-28-2006, 07:46 AM
Lightman

Quote:

Originally Posted by LOE

so all you smart people, can someone explain to me what is the point of 4x4 if it is going to require server mobo, server chip and server memory?

it wont require server memory. Any of you can use some logic and deduct that 4x4 on server board wouldn't take any time to go to market. But designing 4x4 took some time which means that AMD was changing, developing something different than normal 2p server. It certainly will be on 1207 but it won't be re branded 2p workstation board with pcie 16x.
09-28-2006, 09:00 AM
nn_step

Quote:

Originally Posted by LOE

so all you smart people, can someone explain to me what is the point of 4x4 if it is going to require server mobo, server chip and server memory?

I think we reasonably deduced that it isn't really a server Mobo, it doesn't require an exotic server chip or Server ram. So one can assume saying it is on one socket or another is speculation
09-28-2006, 09:02 AM
uOpt

Quote:

Originally Posted by Poodle

Hey guys. It would be pointless of introducing 4X4 with anything other than normal unreg/Non ECC memory modules because if they wouldn't it would be just a normal Socket F system.

Yeah, but apparently AMD tried to do that initially, when they announced 2+2(4x4) to be AM2 based and failed to secure mainboard makers for what is essentially an entirely new platform.

So now that have to share with an existing platform.

Quote:

Last year I believe that Abit showed of an Opteron socket 940 (e.g. not 939) board that didn't make it to the shops but seemed to work, that could handle unreg/non ECC. So why wouldn't all the big guys over at Amd not be able to get this to work?

Oh my, not that rubbish again. You are not the troll over from 2cpu.com, are you?

That board never made retail and I know nobody who actually ever used one.

Technically, it might be possible. ECC is never required and you can move the buffer chips from the RAM modules to the board. Of course, that would be pointless since although you can now use unregistered RAM you still have the register chip limiting overclocks, not to mention it is now more expensive (since you pay for 4 register chips even if you only use 2 modules).

Quote:

Amd wouldn't sell any 4x4 if they wouldn't. We know it and they must surely know it! If i'm wrong the 4x4 was just a media stunt to get people not to jump on Kentsfield.

The cause of events makes it most likely that they tried to establish an AM2 based 4x4 platform with unregistered RAM and failed, and now instead of cancelling 4x4 they just "rebase" it. That means the only technical effect of 4x4 will be that we get 1207 chips with unlocked multipliers and maybe a few boards with overvolting options.

Now, whether anybody produces 4x4 boards (no matter which socket) with serious overvolting options is an entirely different matter. The length of lines on the PCB board and the fact that you have to shuffe around twice the power makes that a very serious risk for the vendor. If you can sell only a couple hundred boards and half of them come back with blown up mosfets you won't make any profit.

In any case, socket 1207 is a registered RAM platform and nobody can do anything about that.
09-28-2006, 09:27 AM
Poodle

Quote:

Originally Posted by uOpt

Yeah, but apparently AMD tried to do that initially, when they announced 2+2(4x4) to be AM2 based and failed to secure mainboard makers for what is essentially an entirely new platform.

So now that have to share with an existing platform.

Oh my, not that rubbish again. You are not the troll over from 2cpu.com, are you?

That board never made retail and I know nobody who actually ever used one.

Technically, it might be possible. ECC is never required and you can move the buffer chips from the RAM modules to the board. Of course, that would be pointless since although you can now use unregistered RAM you still have the register chip limiting overclocks, not to mention it is now more expensive (since you pay for 4 register chips even if you only use 2 modules).

The cause of events makes it most likely that they tried to establish an AM2 based 4x4 platform with unregistered RAM and failed, and now instead of cancelling 4x4 they just "rebase" it. That means the only technical effect of 4x4 will be that we get 1207 chips with unlocked multipliers and maybe a few boards with overvolting options.

Now, whether anybody produces 4x4 boards (no matter which socket) with serious overvolting options is an entirely different matter. The length of lines on the PCB board and the fact that you have to shuffe around twice the power makes that a very serious risk for the vendor. If you can sell only a couple hundred boards and half of them come back with blown up mosfets you won't make any profit.

In any case, socket 1207 is a registered RAM platform and nobody can do anything about that.

1: First I don't hang out much over at 2cpu.com more than to :slobber: when I've heard something new is going on. I don't know what "troll" you're talking about. I don't even got an account there if you want to know...

2. You seem to be sure on your argument so I have to believe you. This said I can't say that 4x4 feels that hot to me anymore . Non ecc/reg was half the point with it. :( I was first glad about it ending up on 1207 but now I don't know...

3. What made me sure of my argument was that the new M2R32-MVP can handle ECC/reg on Am2 so I guessed that the otherway around also was possible.
09-28-2006, 09:42 AM
oldblue

Quote:

Originally Posted by uOpt

Read the disription of the pin layout. The socket 1207 platform is for registered RAM.

Could you point me to a description of the socket F (1207) pin layout? I haven't been able to find one.
09-28-2006, 10:16 AM
uOpt

Quote:

Originally Posted by Poodle

3. What made me sure of my argument was that the new M2R32-MVP can handle ECC/reg on Am2 so I guessed that the otherway around also was possible.

Where's that info coming from? Any pointers?

%%

oldblue, since the memory controller is on the chip the lanes must match the memory access interface. Since we already know that 1207 is for registered with the normal Opterons, there is no way to use the same pin layout for unregistered RAM.

Unless they are willing to re-define what each pin does but then the result would be a new socket, even though it might have the same number of pins. That would instantly kill mainboard maker's interest since you couldn't exchange processors and boards with the other socket 120, that's out of question. Intel got away with doing that with the two 478 sockets, but both of them were major platforms so mainboard vendors could afford it.
09-28-2006, 10:21 AM
Poodle

Quote:

Originally Posted by uOpt

Where's that info coming from? Any pointers?

%%

oldblue, since the memory controller is on the chip the lanes must match the memory access interface. Since we already know that 1207 is for registered with the normal Opterons, there is no way to use the same pin layout for unregistered RAM.

Unless they are willing to re-define what each pin does but then the result would be a new socket, even though it might have the same number of pins. That would instantly kill mainboard maker's interest since you couldn't exchange processors and boards with the other socket 120, that's out of question. Intel got away with doing that with the two 478 sockets, but both of them were major platforms so mainboard vendors could afford it.

Well here: http://www.sweclockers.com/nyhet/?id=4693 in Swedish I'm arfaid. Now I'm not sure anymore if the author was completly sober when he wrote it though. :)

Edit: Maybe he wasn't drunk after all: http://www.neoseeker.com/Hardware/Pr...sus_m2r32_mvp/

Show 100 post(s) from this thread on one page