As far as I know, nowadays dimms are sold in sets. If you want 4 dimms you would buy a set of 4. Two sets of 2 may or may not work and it they boot at standard speeds they will most likely not work at higher clocks.
Using the latest greatest BIOS? Thatâs a must. Do both kits have EXPO? Iâd recommend setting EXPO so it applies the profile settings, but manually reduce the clocks to 4800 before saving.
If the kits donât have EXPO then itâs probably a problem with drive strength impedance/resistance settings, but thatâs something youâd have to pull out your hair to manually tune. I had a hell of a time getting my 96GB kit stable above 5600 until I found the magic drive strength setting combination, but my kit was XMP. Itâs why you should only ever use EXPO enabled memory on AMD systems.
3600 is the supported speed for 2DPC 2R. If ChatGPT was actually useful itâd refer you AMDâs specs. ÂŻ\_(ă)_/ÂŻ
Mmm, this gets repeated a lot but, absent insider info from Asus, AMD or such, itâs not apparent channels get trained individually. And, if they do, it certainly doesnât seem thereâs a BIOS which supports different terminations between A and B. So, in the likely chance inter-kit variabilityâs higher than intra-kit variability, crossing kits on the channels has a reasonable chance at improving matching and yielding a higher stable overclock.
Iâve actually yet to hit a draw on the parts lottery where kit per channel outperforms crossed kits. Doesnât seem to be uncommon the minority of folks who actually try crossing reach similar findings.
That is essencially âworking as intendedâ. It boots fine at 3600 which is the default speed for a fully loaded config. Also, youâre missed Processor info, but if you system was already 2 years old I bet it is a Ryzen 7xxx, which have weaker IMC than 9xxx.
Maybe, maybe not. While Granite Ridge bumps 1DPC to 5600 from Raphaelâs 5200 both support 2DPC 3600. Thereâs a notable lack of evidence to show the prospective 1DPC change has any effect on 2DPC overclocks. Itâs also unclear if AMD actually changed the IO die versus unlocking existing capability or 5600 resulting from backportable microcode or AGESA improvements. If AMD did change the die itâs unclear if Raphael is still being produced with the older stepping or if newer parts get the same step as Granite Ridge.
note that for higher ddr5 speeds this was amended after 2022 to say that there is a speed impact with 1 dimm in 2dpc designs. And âreal-worldâ you can cross kits in the channel to not a negative outcome, but thats not the most tested/qualified use case if the spec is to be believed.
Cool but, hmm, Intelâs definition of mixed differs from crossed as used in this thread. Emphasis added,
Like just about all the other quad DIMM threads around here, this threadâs context is the two kits are the same part number.
In a cross the patterns with generally better odds of better matching between the âidenticalâ DIMMs are
DIMM socket
cross pattern 1
cross pattern 2
A1
kit 1
kit 2
A2
kit 2
kit 1
B1
kit 1
kit 2
B2
kit 2
kit 1
as these nominally try to put the same loading on each arm of each channel. FWIW I havenât yet hit a measurable performance difference between the two cross patterns but I have encountered differences like crossed configs posting 400 MT/s above where the same DIMMs will uncrossed. So far I havenât had crossed be worse than uncrossed and a majority of the time there hasnât been a difference. Only n = 3 for testing across kits, mobos, and Granite Ridge CPUs, though.
Iâve come across a handful of similar results from others scattered around various DDR5 overclock threads, though if anyone else on L1âs tried I havenât noticed it being mentioned.
Iâm having similar issues with my x670e master and 2 kits of identical ram. I watched a video and they recommended putting one kit in all in A the other all in B. They are 7800 single rank stick of 16gb.
My case Iâm aiming for 6000 to 5800 speeds. I think the termination values are what is important not the voltages. Iâm just not sure what ones to try.
So Iâve had the contra experience that two âidentical kitsâ work better in the same channel vs cross channel. Kit1 in channel A and kit2 in channel B have consistently been more reliable for me, outside jedec trays of memory. Even then âno oneâ mixes brands of memory, or even batches that Iâve found, at scale. (Even though the jedec timings and sub timings are exactly the same).
Can you explain your reasoning about intra- and inter- kit variability in that given if you have different part numbers that would be greater variability between two distinct part numbers.
2DPC with Mixed DIMMs means two different DIMM part numbers are populated in each memory slot within a single 2DPC channel (see bottom left image).
2DPC with âidentical part numberâ dimms is also an unsafe assumption imho. You have 4 dimms that have the same part numbers but wildly different ages â mixing the ones purchased at different times have a higher likelihood of working? That certainly hasnât been my experience, at all, and when Iâve seen that working itâs because one dimm in the pair happens to work at looser/more relaxed timings than the others. i.e. older kit wants looser timings which happens to work fine on newer kit, but the other way around doesnât work at all.
Compound that with the fact that the same part numbers can exist with different configurations â even the same chip lots/batches can be far apart â not good advice for someone in search of stability vs kit-per-channel having physically more knobs to achieve signal path integrity and, theoretically, reduced variables.
I donât think youâd disagree that, at a fundamental level, dimms in the same kit should have less variability than dimms from different kits, especially given we can come up with specific examples of wide variability for a given part # (in terms of lots and batches and mfg part #s for constituent chips â but also egregious violators like neemix and corsair where one cannot even count on memory chip brand).
If the goal is to give the system the best shot at dealing with that variability, and the system has more knobs and levers to deal with differences between channel A and channel B than it does knobs and levers to deal with variability between dimm 1 and dimm 2 given the spects⌠Iâm drawn to just one conclusion.
DIMM socket
cross pattern 1
cross pattern 2
A1
kit 1
kit 2
A2
kit 1
kit 2
B1
kit 2
kit 1
B2
kit 2
kit 1
^ This should be someoneâs first goto imho and then only try something else if that fails. This is most likely to be stable in my experience.
This was around n=28 during the 4 dimm testing on both am5 and core ultra â we really put a lot of work testing 4 dimms for 128 and 192gb configurations with mixed and same part numbers. Neemix was by far the worst about same-part-number kits being wildly different, followed by corsair.
I found that motherboard training algorithms would sometimes mask the reality of things not quite working well as they would loosen some of the secondary timings before dropping to a low speed.
Yes, and thatâs probably the main reason why crossing can produce closer matching between channels. I start uncrossed like youâre suggesting and my experience with crossing Vengeance is consistent with you experiences of Corsair variability. Nemix has never been on my buy list because your findings there are entirely what I expected from their business model.
I think the way Iâd put it is if planning to try overclocking 2DPC itâs probably good to buy a quad kit if you can. But quadsâ limited availability is whatâs lead to 2x2x48 configs being common with DDR5 UDIMMs. And 4x48 or 4x64âs an expensive way to go for 1DPC to 2PC upgrades since youâre rebuying two DIMMs. So itâs not the most practical advice. One possible interpretation of the situation is producing effectively matched DDR5 quads is hard enough manufacturers mostly avoid it.
How did G.Skill, V-Color, and Crucial compare for matching between kits?
Can also happen when buying two âidenticalâ two DIMM kits at the same time. Just depends whatâs in the distribution channel and how parts get pulled to box and ship.
Is there a typo here?
If Iâm understanding the direction correctly, the general observation Iâd make is the higher the clock the more important signal integrity matching becomes, meaning matching timing capabilities declines in relative importance. While both sets of concerns are most effectively mitigated by training channels separately to uncrossed two DIMM kits, that doesnât seem to be something the industryâs capable to implement. Rather unfortunate as the available evidence runs almost entirely to DDR5 2DPC being signal integrity bound.
So quad kits seem something of a hardware workaround for software limitations. 4R CUDIMMs another.
My null hypothesis is Adataâs basically rebadging RDIMMs as CUDIMMs by allowing 5 V and moving the notch. The press release not mentioning AM5 at all might be a hint they donât work in bypass mode. Itâs also possible to read the press release as saying only 1SPC boards are supported.
Well, kind of. Itâs pretty vague, the understanding of signal integrity isnât really correct, and the memtest stability check probably doesnât mean much. But itâs better than average out of videos Iâve looked through.
IMO more notably, several overclock attempts are reported in the comments, obtaining speeds of
3600 in three cases
4000 in two
4400 once
with nobody reporting 4800+ that I saw. So itâs sort of an interesting social example of other people looking at the title and then posting alongside about how theyâre expecting to get 5600 or 6000.
2 systems ago I got a 32 GB kit (2 dimms). Then I decided to bump it to 64. Got the second kit and when they did not run at whatever advertised speeds I just returned the second kit. Since then I just get as much memory as possible (or I can afford) in one kit when I build the system and leave it alone. I can certainly relate to the frustration.
EDIT: Right now you can get a 2 dimm kit of DDR5 6400 totaling 128GB for $700-800
Anyone have any ideas for stability? I have tried auto voltages, higher voltages, turn the iGPU off, mess with terminations no matter what I do I canât get above 4800 with 16gbx4 A die.
I also got the impression these might be designed to only work in 1DPC as well, even though they didnât explicitly say it⌠still might be a boon for people with 1SPC boards that want to run 256GB at as high of a frequency as possible.
