News Micron announces its first DDR4 modules

[brumgrunt]

Micron has announced sampling of its first DDR4 modules

http://www.bit-tech.net/news/hardwar.../micron-ddr4/1

[SpAceman]

I hope early adopters aren't going to have to deal with those stupid motherboards with support for multiple memory types. I hated those DDR2/DDR3 combo boards.

[r3loaded]

We'll have a clean break in terms of memory support since the memory controllers are integrated into CPUs. From current rumours, Haswell will still be DDR3, but the next tock Skylake will support DDR4. It still has some way to go before coming to the market in significant numbers.

[schmidtbag]

Originally Posted by SpAceman

I hope early adopters aren't going to have to deal with those stupid motherboards with support for multiple memory types. I hated those DDR2/DDR3 combo boards.

Why? Like sure in the long run they're a horrible choice, but they're still convenient. Much like those AMD motherboard that supported either 754 or 939.



Regardless, I could have sworn I read that DDR4 was going to be skipped and companies were going to move straight to DDR5. We have the technology to do that and its GOING to happen, so why not just skip it? Either way, I'm beginning to feel more and more uncomfortable about RAM these days because latencies are getting to be hideously bad. I realize that isn't as important today as it was several years ago, but I feel like there's a point where there'll be something like DDR8 and it turns out to be slower than previous generations, even though it'd be theoretically much faster speed.

What I'd personally prefer to see is on desktops, ditch these large DIMMs for the smaller SO-DIMMs, with more memory channels. SO-DIMMs generally have the same capacity limits as DIMMs yet they're so much smaller, so you could fit so many more of them on the same motherboard. So think of it like this:
8 DDR3 2GB SO-DIMMs in quad-channel
OR
4 DDR4 4GB SO-DIMMs in dual-channel
I personally feel the first setup would heavily outperform the 2nd, and it'd probably be cheaper.

[r3loaded]

Originally Posted by schmidtbag

Regardless, I could have sworn I read that DDR4 was going to be skipped and companies were going to move straight to DDR5.

You were probably thinking of GDDR4/5 memory, which are both based on DDR3 but with low-level modifications to suit the parallel, high-bandwidth nature of GPUs. Only a few Radeon 2000/3000 series cards used GDDR4 iirc.

[yougotkicked]

@ schmidtbag: I dont' know about converting to SO-DIMMs, but I definitely like the idea of moving to a quad-channel standard. I could imagine a design standard where full-size DIMMs are still used, but each module has two discrete memory banks communicating over separate channels. this way a single physical module would communicate at dual channel speeds, so the basic 2 module arrangement would deliver quad-channel bandwidth. This would theoretically be easier than shifting to SO-DIMMs (mobo makers would not have to change as much, and each module would be able to use the same power circuitry) while opening up quad-channel memory performance to SFF systems. With the shift to 30nm parts it should be more than possible to fit all the needed parts onto a single DIMM (my samsung 30nm DDR3 DIMMs are ~1/2 standard height once you ignore the contact plates for both).

[schmidtbag]

Originally Posted by yougotkicked

@ schmidtbag: I dont' know about converting to SO-DIMMs, but I definitely like the idea of moving to a quad-channel standard. I could imagine a design standard where full-size DIMMs are still used, but each module has two discrete memory banks communicating over separate channels. this way a single physical module would communicate at dual channel speeds, so the basic 2 module arrangement would deliver quad-channel bandwidth. This would theoretically be easier than shifting to SO-DIMMs (mobo makers would not have to change as much, and each module would be able to use the same power circuitry) while opening up quad-channel memory performance to SFF systems. With the shift to 30nm parts it should be more than possible to fit all the needed parts onto a single DIMM (my samsung 30nm DDR3 DIMMs are ~1/2 standard height once you ignore the contact plates for both).

Actually I like your idea a lot better haha, but I don't see how switching to SO-DIMMs would be any more difficult than switching from DDR3 to DDR4. I would think that although the sockets are physically the same size and probably serve similar purposes, I doubt they work similarly electronically.

[mo2580]

Originally Posted by schmidtbag

Why? Like sure in the long run they're a horrible choice, but they're still convenient. Much like those AMD motherboard that supported either 754 or 939.

.

these amd motherboards did they have 2 seperate sockets or was it 1 socket and both processors could be installed on the same socket?

[schmidtbag]

Originally Posted by mo2580

these amd motherboards did they have 2 seperate sockets or was it 1 socket and both processors could be installed on the same socket?

They usually had 1 socket by default, typically 754, and since there wasn't room to fit a 2nd socket, they had a custom-made slot (just above the AGP/PCI-e slot) that looked much like an AGP slot where you could buy this external card that allowed you to mount a socket 939 CPU. Then you had a bunch of jumpers to change so the mobo would know which CPU socket to use. I know asrock was one of the companies that did this but maybe another one did. These motherboards were surprisingly cheap too. You can find these boards on ebay, but I've never seen where you can get the 939 expansion card. I never bothered getting one of these.

[fluxtatic]

Originally Posted by schmidtbag

They usually had 1 socket by default, typically 754, and since there wasn't room to fit a 2nd socket, they had a custom-made slot (just above the AGP/PCI-e slot) that looked much like an AGP slot where you could buy this external card that allowed you to mount a socket 939 CPU. Then you had a bunch of jumpers to change so the mobo would know which CPU socket to use. I know asrock was one of the companies that did this but maybe another one did. These motherboards were surprisingly cheap too. You can find these boards on ebay, but I've never seen where you can get the 939 expansion card. I never bothered getting one of these.

I missed those entirely...and now I kinda want one just because it's bizarre. But only if I could find the 939 card, too.

And is it just me, or has the DDR4 spec taken freakin forever?

[schmidtbag]

Originally Posted by fluxtatic

I missed those entirely...and now I kinda want one just because it's bizarre. But only if I could find the 939 card, too.

And is it just me, or has the DDR4 spec taken freakin forever?

Yup they're very bizarre, especially if you see the 939 card in use.

I feel like DDR2 survived longer than DDR4, there's still new-ish laptops out there that use it. But anyways, DDR4 is kinda pointless, I doubt there will be a significant performance improvement when using it.

[SpAceman]

Originally Posted by schmidtbag

Why? Like sure in the long run they're a horrible choice, but they're still convenient.

Because redundant RAM slots make me angry. When I was looking for a decent spec DDR2 motherboard (back when DDR3 was new and crazy expensive) it seemed like every single board I came across with specs and pricing I liked had support for both types and therefore redundant slots either way. Lucky for me I managed to get myself the wonderful board in my sig.

[yougotkicked]

Originally Posted by schmidtbag

Actually I like your idea a lot better haha, but I don't see how switching to SO-DIMMs would be any more difficult than switching from DDR3 to DDR4. I would think that although the sockets are physically the same size and probably serve similar purposes, I doubt they work similarly electronically.

I was thinking mostly in terms of Manufacturers having to change board layouts. Even though the circuitry probably changes substantially gen to gen, Manufacturers are used to working with the general dimensions and orientation of full size DIMMs. Switching standards would require them to learn how to fit SO-DIMM slots on their boards, which could require making a new standard connector for SO-DIMMs (I don't think there's a vertical oriented slot made for them, heat spreaders wouldn't fit the slanted mount they use for laptops, and RAM makers have sternly refused to stop using spreaders despite nobody who knows anything wanting them).


If the 30nm DDR3 parts I use are anything to go by, the die shrink should allow for speeds upwards of 2500Mhz with ease. We won't know until the DDR4 standard is official, but we could see a pretty major boost in transfer speeds. Hopefully this standard will include more overhead than DDR3 did, I don't want a precedent to be set where we get a new RAM standard every 1.5 generations of CPU architecture, makes things more complicated than they need to be.

[DDR4standard]

LRDIMMs at end-of-life and DDR4 copying NLST IP

The HP Gen8 DL360p and DL380p have support for Netlist (NLST) HyperCloud memory.

This is the memory DDR4 is going to copy (license).

LRDIMMs are already copying it - but have implemented it with asymmetrical lines - leading to latency issues - and inability to deliver 1333MHz at 3 DPC.

With HP Smart Memory HyperCloud - you can get 1333MHz at 3 DPC at 24DIMMs on the 2-processor DL360p and DL380p servers.


LRDIMM buffer chipsets are only made by Inphi (IDTI and Texas Instruments have delayed and backed off respectively). Inphi was the most aggressive against Netlist (NLST) - however they have recently failed in their challenge of NLST IP in patent reexaminations at the USPTO - NLST IP '537 and '274 have been reaffirmed by the USPTO with ALL claims intact - which reflects poorly on Inphi in NLST vs. Inphi.

LRDIMMs thus face legal risk (of possible recall ?).

DDR4 also copies NLST IP - but have chosen to also copy the symmetrical lines and distributed buffer approach.

JEDEC should eventually license NLST IP for DDR4 prior to finalization.

[DDR4standard]

LRDIMMs at end-of-life and DDR4 copying NLST IP

The HP Gen8 DL360p and DL380p have support for Netlist (NLST) HyperCloud memory.

This is the memory DDR4 is going to copy (license).

LRDIMMs are already copying it - but have implemented it with asymmetrical lines - leading to latency issues - and inability to deliver 1333MHz at 3 DPC.

With HP Smart Memory HyperCloud - you can get 1333MHz at 3 DPC at 24DIMMs on the 2-processor DL360p and DL380p servers.


LRDIMM buffer chipsets are only made by Inphi (IDTI and Texas Instruments have delayed and backed off respectively). Inphi was the most aggressive against Netlist (NLST) - however they have recently failed in their challenge of NLST IP in patent reexaminations at the USPTO - NLST IP '537 and '274 have been reaffirmed by the USPTO with ALL claims intact - which reflects poorly on Inphi in NLST vs. Inphi.

LRDIMMs thus face legal risk (of possible recall ?).

DDR4 also copies NLST IP - but have chosen to also copy the symmetrical lines and distributed buffer approach.

JEDEC should eventually license NLST IP for DDR4 prior to finalization.

[noizdaemon666]

Originally Posted by DDR4standard

...snip...

Originally Posted by DDR4standard

...snip...

Double post. And whaaaaaa......? lol

[DDR4standard]

Sorry for double posting.

Basically LRDIMMs and DDR4 are both going to be using NLST IP (search for the article “Netlist puffs HyperCloud DDR3 memory to 32GB – DDR4 spec copies homework”).

LRDIMMs are currently only being made by Inphi, have high latency issues and under-perform as well. So they have failed in copying someone else's IP (no wonder - Inphi hired former MetaRAM CEO as "Technical Advisor" who some years back ALSO conceded to Netlist and went out of business), plus have been unable to copy correctly.

DDR4 does a better job - as they are doing a better job of copying the Netlist IP (expect JEDEC to license prior to DDR4 finalization).

[DDR4standard]

The technology in LRDIMMs/NLST HyperCloud is called "load reduction" and "rank multiplication".

Currently you only need it at 3 DPC with Romley (as 16GB RDIMMs 2-rank are cheap and work well at 1 DPC and 2 DPC).

3 DPC usually used for virtualization/data centers (lots of memory per server for all the VMs you want to run).


When 32GB RDIMMs arrive, however the need for this technology will shift to 2 DPC also (i.e. more mainstream even) - because 32GB RDIMMs will be 4-rank (as 2-rank cannot be made for a couple of years because of lack of 8Gbit DRAM die).


Both LRDIMMs and NLST HyperCloud are available from IBM and HP. HyperCloud is called IBM HCDIMM and HP HDIMM (HP Smart Memory HyperCloud) - delivers 1333MHz at 3 DPC. HP ships it with their virtualization servers DL360p and DL380p only at 3 DPC (i.e. fully loaded 24 DIMMs per server).