Last week the OS/2 Museum got its first ever set of real (i.e. not fake) 4 GB DDR2 desktop (unbuffered) DIMMs, a pair of Samsung 4GB PC2-6400U modules. Such modules are quite rare and correspondingly tend to be unavailable at a reasonable price. Moreover there is a lot of fake “AMD only” modules. But why are the real DDR2 4 GB UDIMMs so rare?
The short answer is “because they’re almost useless”.
The long answer is that DDR2 modules based on 2 Gbit memory technology became available too late in the DDR2 life cycle, which led to such modules being rather poorly supported. They are not unsupported—there are DDR2 boards which not only accidentally happen to work with such modules but where the manufacturers officially support them.
A case in point is the Intel DQ45CB board. It came out in mid-2008 and when it was released, it did not list 4 GB modules as supported, most likely because they weren’t available. The board has four memory slots and it was initially specified to support up to 8 GB memory maximum (with four 2 GB modules). Sometime between April 2009 (specification update E53961-004US) and May 2010 (specification update E53961-006US), Intel officially changed the board’s specification to support up to 16 GB RAM when using 4 GB DIMMs based on 2 Gbit memory technology.
Obviously at that time DDR3 RAM was already available. In fact the same Intel Q45 chipset supports both DDR2 and DDR3 (though any given board would naturally only support one or the other type). This led to some curious situations; the above mentioned Intel DQ45CB board eventually supported up to 16 GB DDR2 memory, but the “better” DP45SG board from the same generation used DDR3 memory and never (at least officially) went beyond 8 GB maximum memory support.
It is worth noting that even though it was not officially listed as supported at the time, a DQ45CB board with BIOS version 0059 dated 08/25/2008 works just fine with the 4 GB modules. There is nothing in the DQ45CB BIOS revision history mentioning that support for 4 GB DIMMs was added to the firmware at some point after the initial release.
For good measure, I tried the 4 GB DIMMs in an older Intel D975XBX2 board. To my surprise, the board actually managed to POST, but the BIOS was very confused about the amount of installed memory and memtest86 crashed and burned almost instantly. Not that it was expected to work.
As is often the case with the largest memory modules of any given generation, there are several separate problems. The memory controller needs to support the given memory technology and DIMM architecture. Without that, the DIMMs can’t work at all.
On the Intel side, I believe that only the 4 Series chipsets (P45, Q45, P43 etc.) supported 4 GB DDR2 modules (2 Gbit technology). That was also Intel’s last generation with DDR2 support and only came out a few months before the next-generation X58 chipset was released. In fact the 4 Series chipsets also supported DDR3, not just DDR2. Intel’s 3 Series and all older DDR2 chipsets did not support 2 Gbit DDR2 chips (although even the 3 Series chipset supported DDR3 already).
A separate problem becomes relevant when one tries to fully populate the memory. It might happen that, say, the chipset supports 4 GB modules but only supports up to 8 GB physical memory, and still can’t use 16 GB with four memory slots populated. I don’t believe this was the case with the Intel desktop chipsets and DDR2 4 GB modules; the 4 Series chips could support 16 GB physical memory, and the older ones with 8 GB maximum physical memory support didn’t work with 2 Gbit memory technology anyway.
Finally even when the chipset supports the memory technology and can address fully populated memory, the firmware needs to properly support it. Sometimes the hardware is capable but firmware gets in the way. That happened at several points with various memory generations and boards.
In summary, the reason for the rarity of 4 GB DDR2 modules is that by the time the chipsets with 4 GB DDR2 DIMM support were released, DDR2 was almost obsolete. The situation was different in the laptop market and that is why 4 GB DDR2 SO-DIMMs are not quite as exotic.
Things were somewhat different in the AMD camp. Since the memory controller was on the CPU, the board didn’t fully determine what memory was supported. Technologically there was another major difference between AMD and Intel: In the DDR2 times, Intel’s server hardware almost exclusively (the 5100 San Clemente chipset being the exception) used FB-DIMMs (fully buffered DIMMs), modules that looked similar enough but were significantly different from desktop memory. AMD’s Opterons on the other hand always used registered DDR2 which was not vastly different from typical desktop RAM.
High Density vs. Low Density, Ugh
For a long time now, there’s been debate about “high density” vs. “low density” DDR2 RAM and what is or isn’t supported by Intel vs. AMD. As it often is with Internet debates, the signal to noise ratio is extremely low.
The word on the street was that AMD CPUs, including desktop CPUs, supported “high density” DDR2 modules and Intel did not. The truth is that AMD desktop CPUs can work with non-standard 4 GB (and 2 GB) memory modules that do not work in Intel systems. Some of that has been previously discussed on this site.
As a case in point, let’s do a comparison. These are my 4 GB modules, Samsung M378T5263AZ3-CF7:
Here’s a product photo (from an eBay auction) of a module with the exact same Samsung M378T5263AZ3-CF7 designation but obviously a very different layout; in fact while mine are labeled as 2Rx8 (dual rank, x8 chips), the other set is labeled 2Rx4:
According to Samsung, a module with M378T5263AZ3-CF7 designation has sixteen x8 chips (each chip with 256 MB or 2 Gbit capacity), from which one can conclude that mine are actual Samsung DIMMs and the others are one of the many “AMD only” fakes.
In fact there is another indication that the module is fake. That “4GB 2Rx4 PC2-6400U-666-12-E3” designation is actually a standard JEDEC product label. The last “E3” bit refers to a reference design using third revision raw card ‘E’ as given in the JEDEC standard. And, of course, raw card ‘E’ uses a 2Rx8 architecture with 16 chips total, not 2Rx4 with 32 chips total.
But back to density. No one knows what “high density” memory really means, because the term is by definition relative (and has been used with pre-DDR SDRAM, for example). Maybe “high density” refers to DDR2 memory with 2 Gbit chips, although sometimes it does not. As we’ve seen above, there are some Intel boards that actually do support such memory, even though they’re uncommon.
There appears to be some agreement that “high density” refers to the capacity of individual memory chips, not the number of chips on a DIMM. But then we have guides like this one where we’re told that 1 GBit chips are high density, but 1 Gbit chips are also low density (see explanation below on what the real problem is). To be fair, others have made the same claim.
There are also nuggets like this: “The easiest way to tell if your DIMM is high density is if there’s no name or manufacturer on the DIMM itself”. So now we know why unscrupulous sellers bother slapping obviously fake Samsung labels on sketchy DIMMs: That makes them low-density and therefore good!
Then we have Crucial, an actual memory manufacturer, telling us pretty clearly that “high density” means higher capacity chips, not that one 1 Gbit chip can be “low density” and another 1 Gbit chips “high density”.
But it’s even worse than that. That fake 4 GB Samsung module above obviously has 1 Gbit chips, that’s why it needs so many to get to 4 GB capacity. And that module does not work in an Intel board, whereas the real Samsung DIMM with higher density chips does work. So it’s not that “low density memory works in Intel boards and high density doesn’t”; when it comes to 4 GB DIMMs, only RAM based on high-density chips works in Intel boards!
The bottom line is that the terms “high density” and “low density” in the context of DDR2 RAM have been abused to death, are effectively meaningless, and should be avoided.
What Makes UDIMMs AMD Only?
OK, so if it’s not “density”, there must be something else that makes those “AMD only” DIMMs incompatible with Intel hardware. The answer is in the actual DDR2 UDIMM standard. Remember that “2Rx4” designation on the fake Samsung modules? That’s probably the only part of the label that isn’t fake.
According to the JEDEC PC2-5300/PC2-6400 DDR2 SDRAM Unbuffered DIMM Design Specification (Revision 3.2, 2012) there simply are no UDIMMs based on x4 architecture! Only x8 and x16 based UDIMMs are specified. Registered DDR2 for servers is different and can use x16, x8, and x4 chips (in fact x4 DDR2 server DIMMs are probably the most common). An older version of the DDR2 UDIMM specification can be found here; the 2012 update version can be obtained from the JEDEC website but it’s not very different.
Since Intel used completely different memory technology (FB-DIMMs) in their DDR2 server hardware, their server and desktop/mobile memory controllers were very different and the latter never supported x4 DDR2 memory chips. It’s not clear which was the chicken and which was the egg, but it is apparent that DDR2 UDIMMs never defined x4 modules because Intel never supported them, and Intel never supported x4 UDIMMs because they were never defined by the JEDEC standard. Certainly Intel’s 3 Series and 4 Series datasheets clearly say that they support x8 and x16 memory, while x4 is not mentioned.
Bottom line, DDR2 unbuffered DIMMs with x4 chips are non-standard, fake, and although they might work in AMD systems just fine, they will definitely not work with Intel chipsets.
Conclusion
So are the real 4 GB DDR2 DIMMs any good? Yes, if you have the hardware that can use them, which in the Intel world means a 4 Series chipset. I can imagine that they could be really useful for example in an Intel DQ45EK board which only has two memory slots. And in a DQ45CB board, or in one of the handful of non-Intel brand boards that reportedly support 4 GB DDR2 UDIMMs, it’s a rare opportunity to combine an Intel chipset based desktop Core 2 platform with 16 GB RAM.
You’re reading my mind Michal – I have a DQ45CB here, and was wondering where the hell the /real/ 4 GB DDR2 DIMMs are.
There are now 16GB AMD-only DDR3 UDIMMs on eBay. I guess it’s the same kind of pushing of limits as these 4GB DDR2 were. And by now the vendors have learnt they don’t need to fake anything, so they just look like this: https://i.ebayimg.com/images/g/ihEAAOSwTFJfKn1k/s-l400.jpg
Wonder if my Gigabyte GA-EX38-DS4 supports these mythical 4GB DIMMs. The board is certainly new enough (2008). I know this board has been kneecapped by Gigabyte to not support things like VT-d. At the time I went with a DDR2 board since DDR3 memory was so damn expensive.
The X38 datasheet says: “Supports 1-Gb, 512-Mb DDR2 or DDR3 technologies for x8 and x16 devices” and 8 GB RAM maximum, with nothing relevant in the spec update. So I would say almost certainly not.
That’s actually almost the same as the slightly newer X48 chipset, which is really more like X38+. The only difference in the X48 is the lack of DDR2 support, but the limitations remain (no 2 Gbit chip support, 8 GB max).
The 4 series chipsets was the only one that listed 2Gbit DDR2 in the datasheet, but this does not mean it was unsupported on older chipsets. I believe that the hardware support had existed since the 965 chipset. The same is true for 2Gbit DDR3 too BTW. 2Gbit DDR2 tended to be large die size (as you can notice from the Samsung sticks in the photo) while 2Gbit DDR3 came later with a smaller die size.
The press releases for Samsung 2Gbit DDR2 and 2Gbit DDR3:
https://www.samsung.com/semiconductor/newsroom/news-events/samsung-introduces-60nm-class-processing-for-2gb-ddr2-dram/
https://www.samsung.com/semiconductor/newsroom/news-events/samsung-samples-smallest-2-gigabit-ddr3-with-power-savings-of-40-percent-or-more-for-servers-and-pcs/
And yes there are even X48 motherboards with DDR2 out there, despite what the datasheet say too.
I think this is partly because early 1Gbit DDR3 in 2007 had a quite high die size premium.
I remember the same garbage “high density” thing with SDR DIMMs in the early naughts. They didn’t work (well) on intel boards either.
As a side note, I’m pretty sure that nearly all Intel platforms with integrated memory controllers can support 4Gbit DDR3 in hardware. Intel’s Core 2 DDR3 chipsets don’t have enough address lines for it though.
“There are now 16GB AMD-only DDR3 UDIMMs on eBay. ” No doubt there are similar 2Rx4 registered DIMMs as well. I assume that the LRDIMM buffers are a bit hard to get these days, right?
Intel Nehalem IMCs didn’t handled correctly 4 GBit DDR3 DRAM chips. I remember a few Threads on Hardware forums with people managing to POST but having constant stability issues using 8 GB DDR3 modules and late Single Sided/Single Rank 4 GB ones. Solution was getting old Dual Rank 4 GB modules.
It may be possible that someone manually configuring all the Subtimmings may manage to get them working, since it was promising than they actually booted Windows instead of not POSTing at all. But don’t recall if someone had a success history with them.
Was the problem limited to early processors like Xeon 5500?
Not sure. Actually, I remembered than the users with issues were always using consumer LGA 1156 with Core i3/i5 Lynnfield or Clarkdale, not the LGA 1366 Core i7 Bloomfield. So it may be possible that the 4 GBit DDR3 works in the HEDT platform but not consumer. May be worth to test that in the name of science.
The datasheet for Bloomfield (45nm Core i7 LGA1366) says that DDR3 “512Mb, 1Gb, 2Gb, technologies/densities are supported”, 24 GB RAM maximum. The Lynnfield (45nm Core i7 LGA1156) datasheet says “1-Gb and 2-Gb DDR3 DRAM technologies are supported”, 16 GB RAM maximum. The Xeon 5500 series datasheet makes no clear statement.
The Gulftown datasheet makes no clear statement either but despite Intel’s claim that Gulftown is limited to 24 GB, there are people who successfully ran it with 48 GB RAM (obviously using 8 GB UDIMMs, presumably based on 4 Gbit technology).
Intel’s DX58SO2 Technical Product Specification makes a weird claim that the board supports “up to 24 GB of system memory with 6 DIMMs using 4 Gb memory technology” — that does not add up. Funnily enough Intel’s own product brief for the same board clearly says 48 GB max. If there’s interest, I can dig up my DX58SO2 board and see what it can do with Nehalem/Westmere Core i7 and Xeon CPUs.
It is totally possible and maybe even likely that the LGA1156 machines are limited to 16 GB RAM regardless of the DIMMs used.
Here is a press release:
https://www.samsung.com/semiconductor/newsroom/news-events/samsung-mass-producing-40nm-class-8gb-ddr3-module-for-laptops-and-mobile-workstations/
I’ve got 16GB here in an HP xw4600 (X38 based, DDR2 ECC unbuffered). However, the system I actually bought it for, the Intel S3210SHLC, gets very confused. I only managed to get it to recognise 10GB, and then with issues, 8GB is the most I could manage for reliable operation.
That was a real pity because VT-d works on the S3210SHLC and is really hampered by 8GB. The xw4600’s board should on paper be an almost perfect Core2 system, but the BIOS is quite flawed with PCI-e IRQ assignment and VT-d that doesn’t work at all.
I hate when that happens, perfectly good hardware marred by subpar firmware.
I have old HP Compaq dc 7900 tower that uses motherboard 3032h. It has four banks each with 1G Ram. These banks handle max. 4GB. While searching for DDR2 PC2-6400 non-ecc unbuffered, I always find 4GB with label for AMD only. So I got confused. To Ram or not to Ram. After reading this article, I think that I should be looking for high density chips only. Did I understand that correctly?
Yes. The “AMD only” RAM is labeled that way for a reason, it really doesn’t work with Intel boards.