RAM explained: Why two modules are better than four / single vs. dual-rank / stability testing

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
I wouldn't make CPU OC ability such an important part of the buying decision. AMD have chosen a decent compromise for their voltage-per-frequency curves, leading to high efficiency of their CPUs especially in multithreaded workloads. If you OC the 5950X, you can get single-digit performance improvements, but this will usually come with a cost of 25-50% higher power consumption, and thus higher temperatures and lower efficiency.

But yeah, the X3D with the 3D-stacked cache on top of the chip is sort of a "first experiment" for gamers that want the additional benefits (as games really seem to profit from it), and the cache hasn't been optimized for overclocking at all yet. If you don't have gaming as the main priority, then a normal "X" CPU model would probably be better.
 

The_King

New member
Joined
Feb 5, 2022
Messages
19
I wouldn't make CPU OC ability such an important part of the buying decision. AMD have chosen a decent compromise for their voltage-per-frequency curves, leading to high efficiency of their CPUs especially in multithreaded workloads. If you OC the 5950X, you can get single-digit performance improvements, but this will usually come with a cost of 25-50% higher power consumption, and thus higher temperatures and lower efficiency.

But yeah, the X3D with the 3D-stacked cache on top of the chip is sort of a "first experiment" for gamers that want the additional benefits (as games really seem to profit from it), and the cache hasn't been optimized for overclocking at all yet. If you don't have gaming as the main priority, then a normal "X" CPU model would probably be better.
I didn't go with the 5700X or 5800X3D rather the 5600X paired with some Patriot Vipers 4400 @ 3800 CL-15-15-15. Really very happy with this setup and performance!
Also no WHEA errors at 1933 FCLK.
5600X CPUZ AGESA 1207 691 ST crop.jpg


3800 CL15 1.44v TM5 Pass 5 cycles crop.jpg
3867 CL16 TM5 5 cycles bcrop.jpg
 

Doc_Bucket

Member
Joined
May 12, 2022
Messages
59
I have B460M Mortar WIFI and did some tuning of timings but I was able to run memtest86 only once. Every other time it would not boot from USB (I checked the priority all right). Does anyone have experience with this issue, or an idea why it did so? Is there a setting to change in BIOS?

I ran other tests of course (Prime95 Large FFTs, Memtest 7.0, GSAT) but none of them from outside Windows or with a 100 % RAM usage. I hope it will be enough, but do not know for sure.

I will be obliged for any suggestions.
 

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
For Memtest86, don't change the boot priority in the BIOS, just press F11 during POST for the boot menu, then select the first entry of the USB stick (because there are two boot entries on there).

But Memtest86 is not that important, the other tools will discover RAM instability sooner or later if you let them test for long enough, you can be pretty much sure about it. I use Memtest86 only in cases where i'm not sure about stability at all, like on a new system, with new RAM modules, or the first time i really tune the settings. Meaning, before i even start Windows, because i don't wanna risk any major instability corrupting my Windows installation.

Later on during RAM tuning, you get to know your RAM and its capabilities a bit better, and you get to know how far "on the edge" you are with your settings. Then, Memtest86 won't tell you anything different from the other tools. I don't even tend to run it anymore then, maybe only at the end as a finisher, for peace of mind. I focus on TM5 for quick testing (if that already fails, it's back to the drawing board for the settings), then GSAT, Memtest (Windows), Linpack Xtreme for the more thorough testing. I also add a bit of extra voltage than what i find to be stable, to account for worse circumstances like higher ambient temperatures which can decrease RAM stability. You always wanna leave a bit of headroom.
 

Doc_Bucket

Member
Joined
May 12, 2022
Messages
59
That's great. Thank you for more tips. I was actually surprised how much can be done with timings.
 

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
Yes, there is a large potential in the timings, since anything but the four main primary timings tCL, tRCD, tRP and tRAS tend to be set very conservatively in the XMP profile. People that don't know much about RAM may mainly look at the frequency, like DDR4-3600. People that know a bit more may also look at CAS Latency, like CL16. They might even look at the four primary timings that are usually stated by the memory brands, but they give most attention to tCL, not knowing that tRCD and tRP and their discrepancy to tCL can actually tell you more about the quality of the RAM.

But then, those who really know about RAM know that the XMP profile is still far below what the RAM can actually achieve in regards to some other timings. So they will look at the four primary timings just to gauge the quality of the RAM (which memory ICs were most likely used), but they will know that XMP leaves too much performance on the table, simply because the RAM brands prioritize stability and compatibility when defining those other timings, Which is not a wrong approach, the RAM brands simply want to make their RAM modules work properly in all kinds of boards and configurations. So you will always have to do the fine-tuning at home if you want to maximize the performance.

Tuning the timings is not for everyone, because it is a complicated subject if you want to know what you're doing. It can be very time-consuming not only for the research (which most people tend to skip), but also for the stability testing. And the reward on the overall performance is perhaps not as big as one might expect, after investing so much time into it. So for most casual users, i fully get it when they just enable the XMP profile and are happy with that. That's completely valid. But i invested the time in finding out more about the timings, so i enjoy getting more out of my RAM now.
 

Doc_Bucket

Member
Joined
May 12, 2022
Messages
59
Exactly, you are describing my experience. Originally I thought only about frequency (which I cannot change on B460) and tCL, taking but a passing notice of "those other three numbers". I imagined there wouldn't be much gain beyond them. But tRFC, tCWL and tRTP proved equally influential, as well as tREFI and some other secondary numbers. A few full days later I finished at 95–96,5 % of hypothetical max bandwidth, with latency below 50 ns. tCL and tRSF scale well with voltage in case of Hynix CJR, they say, and I can confirm that. However, there was not much room left for tRCD and tRP. I wondered why BIOS would not allow me to decrease tRAF, while tRCD remained the same and I kept lowering tRTP. This way the equation tRAF (28) = tRTP (6) + tRCD (14) would not be true. I hope tRTP is not too low now, but tests I've mentioned returned no errors.

Finally, I've come to your conclusion: there's always some potential left because XMP must provide sufficient headroom to make sure the given settings will work in any case. Tightening timings does not feel like pushing the RAM beyond its natural limits. It's not about buying a faster stick, which would be pointless in case of a locked MB anyway, but exploring that potential. An interesting entreprise.
 

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
You wrote tRSF and and tRAF where you probably meant tRAS, but i understand. A bunch of formulas you find in DDR4 overclocking guides are not entirely correct. For tRAS, the minimum should be tRCD + tRP (not tRTP). Often times, one or two clock cycles are added. For example, on my dual-rank Samsung B-Die, i use DDR4-3600 tCL14 - tRCD 15 - tRP 15 - tRAS 32, when i could've gone 14-15-15-30 instead of -32, but -30 gave me trouble. For tRTP i use 8, i could only use tRTP 6 on my single-rank B-Die before. tRFC and tREFI are also important for RAM performance, but leave some headroom there, don't go to the lowest tRFC that the board POSTs with for example, this will lead to trouble.

You can study for days and days how all the timings are related. Once you did that (the level i did it to, it really requires some interest in the topic) then you start to spot inconsistencies even in pretty good DDR4 guides like this one, for example some of their "safe/tight/extreme" classifications i don't really agree with, some are too conservative and others are too aggressive. Still a nice resource though. But i didn't look primarily at that kind of guides, instead i looked into the papers and publications that deal with DDR4 specifications and timings. Some people do their dissertation on a related topic or there's a book on it. Those sources actually give more insight than some OC guide which just says how low a timing will theoretically go.
 

Doc_Bucket

Member
Joined
May 12, 2022
Messages
59
:) You've probably guessed that's the guide I used. For my level a great compilation. "tRAF" – all those letters still get mixed up in my mind… But you are right, this is just the beginning based on "trial and fail" method. Hopefully I will delve into the subject deeper, it looks like real alchemy.

In terms of the above mentioned equation, I was referring to a figure linked from the guide as well ( 118769121-298a6000-b8c3-11eb-8793-7d90e885ca67.png (2062×494) (user-images.githubusercontent.com) The image suggests tRAS does not overlap with tRP, is that incorrect?

Practically, I find what you state (i.e. tRAS = tRCD + tRP) is true. Auto settings also suggest a value by 2T higher, yet I keep it at 28 for now. I experienced more than once that BIOS would not let me change some numbers, or if it did, they would appear reset after booting. It makes me feel MSI makes the whole business safer for users.
 

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
I was talking more about the minimum for practical use which i have found to be the case. Because theoretically you can go lower than tRCD + tRP, so i should probably not call this a "formula" either, but it's just that in practice, i found it's not beneficial or gives problems when you go lower. As you said, you made similar observations. So i find this "formula" to be more correct than the other one.

For some timings there are minimums of course, meaning even if you set them lower, the memory controller will ignore your setting and just use the minimum. And a lot of timings are interconnected, so the minimum will be dynamic, i.e. depending on some other timing, or you may actually have to modify the other timing instead to get the initial timing to the value you want. The BIOS may regard those interconnections, which is good.
 

7620772155602dd

New member
Joined
May 3, 2022
Messages
6
Since some people run into problems with four RAM modules on modern MSI mainboards, i wanted to explain the reasons behind that, and why two modules are often superior.
The main reason lies in the way the memory slots are connected to the memory controller, which is inside the CPU. So the first explanation is about:


1) RAM slot layout

All regular mainboards and desktop CPU models have a dual-channel memory system. Since a lot of boards offer four RAM slots, a pair of two slots have to each form a RAM channel.
So the four RAM slots are not individually addressed, but in pairs, as two channels. The different ways to connect the RAM slot pairs on the board are either "Daisy chain" or "T-Topology".
This RAM slot layout decision - the way the slots are connected - has a big influence on how many modules (two or four) the board works best with.

Here is a slide from an MSI presentation, showing that almost all of today's boards have a "daisy chain" memory slot layout. This layout heavily prefers two-module-operation.
The presentation is a bit older, but it's safe to say that the the vast majority of recent mainboards (B550, Z590, Z690 etc...) also have a daisy chain layout, and it's confirmed in several reviews.
Especially MSI are known to use this layout on almost all their modern boards. For other mainboard makers, it depends on the board model, but they will also tend to prefer this layout.

Daisy Chain.jpg


Daisy chain means that the slot pairs are connected one after the other, and therefore optimized for two modules total. The right slot of each channel is the end point.
Using two RAM modules, they are to be inserted into slot 2 and 4 counted from the left as per the mainboard manual. Meaning, into the second slot of each channel and thus the end point.
The reason is, this puts them at the very end of the PCB traces coming from the CPU, which is important for the electrical properties.
PCB (printed circuit board) traces are the thin signal lines that are visible on the mainboard, especially between the CPU and the RAM slots.

View attachment 149843

Why is this important? The PCB traces, going from the memory controller contacts of the CPU, to each contact of the RAM slots, are optimized to result in exactly the same distance between all those points. They are essentially "zig-zagging" across the board for an electrically ideal layout, making a few extra turns if a direct line would lead to an uneven distance.

This is done so that, with two modules, a) each RAM module is at the very end of the electrical traces coming from the CPU's memory controller, and b) each module has exactly the same distance to the memory controller across all contacts. We are dealing with nanosecond-exact timings, so all this matters.

On a mainboard with a daisy-chain RAM slot layout, this optimization is done with only two modules in mind, which are in slot 2 and 4 (on the board, those slots are called A2 and B2).
This is the configuration that most buyers would use, and it also results in the best overclocking potential. This way, the mainboard makers can boast with higher RAM overclocking frequencies when advertising the board, and the majority of buyers will have the ideal solution with two RAM modules.

Note: Never populate slots 1 and 3 first. When putting the modules into slot 1 and 3, the empty slots 2 and 4 would be like having some loose wires hanging from the end of each RAM contact, creating unwanted signal reflections and so on. So with two modules, they need to go into the second slot (slot 2+4, or A2 and B2) of each memory channel, to not have "loose ends" after the RAM module.

View attachment 155049


Now the interesting question. What happens when we populate all four slots on a mainboard with a daisy-chain slot layout? Well, the module in the second and fourth slot become "daisy-chained" after the modules in the first and third slot. This completely worsens the electrical properties of the whole memory system.

With four modules, there are now two modules per channel, and the two pairs of a channel don't have the same distance from the memory controller anymore. That's because the PCB traces go to the first slot, and then over to the second slot. This daisy-chaining - with the signal lines going to the first and then to the second module of a memory channel - introduces a lot of unwanted electrical handicaps when using four modules. The signal quality worsens considerably in this case.

Only with a "T-Topology" slot layout, the PCB traces have exactly the same length across all four slots, which would provide much better properties for four-module operation. But mainboards with T-Topology have gone a bit out of fashion, since most people use just two modules. Plus the memory OC numbers look much better with a daisy chain layout and two modules. So if the mainboard makers were to use T-topology on a board, they couldn't advertise with such high overclocking numbers, and people would think the board is worse (and it actually would be, for only two modules).

View attachment 156014
Example of an ASUS board with the rare T-Topology layout, advertising the fact that it works better with four modules compared to the much more common boards using the daisy-chain layout.


2) Single-rank vs. dual-rank

Another consideration is single-rank vs. dual-rank modules. This is about how a RAM module is organized, meaning, how the individual memory chips on the module are addressed. To put it simply, most (if not all) 8 GB modules are single-rank nowadays, as well as some 16 GB modules. A single-rank module puts less stress on the memory system. There's also a bunch of 16 GB modules that are dual-rank, and all bigger modules are always dual-rank. Dual-rank is slightly faster performance-wise (up to 4%), but also loads the memory controller more. One dual-rank module puts almost as much stress on the memory system as two single-rank modules!

What is the memory system? It consists of the CPU's integrated memory controller (IMC), the mainboard and its BIOS, and the RAM itself.

Every modern mainboard will be the happiest with two single-rank modules (for dual-channel operation), because this causes the least stress on the memory system, and is electrically the most ideal, considering that the memory slots are connected as "daisy chain". This fact is reflected in the maximum DDR frequencies that the mainboards are advertised with.

Here is an example from the highest MSI DDR4 board model using newest Intel Z690 chipset, which should therefore be quite advanced (specs of MPG Z690 EDGE WIFI DDR4, under "Detail"):
Max. overclocking frequency:
1DPC 1R Max speed up to 5200+ MHz
1DPC 2R Max speed up to 4800+ MHz
2DPC 1R Max speed up to 4400+ MHz
2DPC 2R Max speed up to 4000+ MHz

"DPC" means DIMM (=module) per channel, 1R means single-rank, 2R means dual-rank.

With 1DPC 1R = two single-rank modules (so, 2x 8 GB or 2x 16 GB single-rank), the highest frequencies can be reached.
With 1DPC 2R = two dual-rank modules (like 2x 16 GB dual-rank or 2x 32 GB), the maximum attainable frequency is lower, since the memory system is under more stress.
With 2DPC 1R = four single-rank modules (4x 8 GB or 4x 16 GB single-rank), the maximum frequency drops again, because four modules are even more challenging than two dual-rank modules.
And 2DPC 2R = four dual-rank modules (like 4x 16 GB dual-rank or 4x 32 GB) combines the downsides of the highest possible load on the memory controller with the electrical handicap of using four slots on a daisy-chain-mainboard.

The last configuration can already be difficult to get stable at DDR4-3200 sometimes, let alone DDR4-3600. One could consider themselves lucky to get DDR4-3600 working with four dual-rank modules, maybe having to use more relaxed timings for example. The 16 GB and 32 GB modules also often don't have particularly tight XMP timings to begin with, like DDR4-3600 18-22-22-42.
By the way: The second timing (tRCD) is more telling and important than the first one (tCL) to determine the module quality, but most people only look at the first one, CAS Latency.

With the new DDR5 standard, this drop in attainable frequency is even more pronounced. This is from the specs of one of the top MSI Z690 boards (specs of MEG Z690 ACE, under "Detail"):
Max. overclocking frequency:
1DPC 1R Max speed up to 6666+ MHz
1DPC 2R Max speed up to 5600+ MHz
2DPC 1R Max speed up to 4000+ MHz
2DPC 2R Max speed up to 4000+ MHz

When going from two modules (first and second entry) to four modules, the attainable frequency drops drastically. With two single-rank modules, DDR5-6000 and above is possible according to MSI. With two dual-rank modules, that goes down a little already. But with four modules, the memory system is under a lot more stress, and MSI are quite open about the result. This seems to be a limitation of the DDR5 memory system, which relies even more on a very clean signal quality. Using four DDR5 modules on a board with a daisy-chain layout clearly is not good in that regard.
This deterioration with four DDR5 modules is so drastic that the conclusion could be: DDR5 motherboards should come with only 2 dimm slots as standard (Youtube)


Generally, in case of RAM problems, no matter the technology, there are three possibilities, which can also be used in combination:
1) Raise voltage
2) Lower frequency
3) Loosen timings

But in some cases, buying different modules is actually the best solution.


3) Amount of RAM

32 or 64 GB RAM can be justified for large video editing projects, rendering, heavy photoshop use and such cases. But if it's gaming for example, modern games very rarely use more than 16 GB RAM. There are just the first games coming out, like Flight Simulator 2020, who will use a little more than 16 GB RAM, but still basically run at the same speed. So, 32 GB would definitely be a generous amount, but it can be befitting for a high-end gaming system. However, 64 GB amounts to a waste of money for gaming, no matter what. Before any game will ever touch more than 32 GB, the whole PC will be long outdated, because it could take years. Why are games satisfied with 16 GB nowadays? Consoles. A lot of games are developed with the lucrative console market in mind, and even the PlayStation 5 only has 16 GB of RAM. So games are designed from the ground up not to need more RAM, which then also applies to the PC versions of those games.

Recommendations for use on modern consumer mainboards:
8 GB RAM: Use 2x 4 GB, or even 1x 8 GB if RAM performance isn't critical anyway (entry-level systems, office work etc.)
16 GB RAM: Use 2x 8 GB (good mid-range / gaming systems, most of the time you won't need more)
32 GB RAM: Use 2x 16 GB (high-end gaming systems - when all other bottlenecks are removed - and uses beyond gaming, such as video processing etc.)
64 GB RAM: Use 2x 32 GB (uses beyond gaming), and consider using a professional quad-channel-memory platform like Threadripper with 4x 16 GB.
128 GB RAM total or higher: Use 4x 32 GB, but now you are maximally stressing the memory system in most configurations, so seriously consider using a professional quad-channel-memory platform like Threadripper with 4x 32 GB or more.

I usually recommend DDR4-3600 frequency with AMD and Intel Rocket Lake 11th gen / Alder Lake 12th gen (see chapter 4).


3b) How to increase the RAM size when you have 2x 4 GB or 2x 8 GB RAM?

First choice: Replace the 2x 4 GB with 2x 8 GB, or the 2x 8 GB with 2x16 GB. The new RAM should be a kit of matched modules. This will ensure the best performance and the least problems, because there's only two modules again in the end.

Second choice: Add a kit of two matching modules to your two existing modules. But you might not be able to get the same modules again. Even if they are the same model, something internally might have changed. Or you might toy with the idea of adding completely different modules (for example, adding 2x 8 GB to your existing 2x 4 GB). This can all cause problems. The least problems can be expected when you add two modules that are identical to your old ones. But then there's still this: You are now stressing the memory system more with four modules instead of two, so the attainable RAM frequency might drop a little. Also, it's electrically worse on a mainboard with daisy-chain layout, as explained under 1).

Lastly, adding just one more module (to have three modules total) is by far the worst choice for several reasons. Every desktop platform has a dual-channel memory setup. This means it works best with two modules, and it can work decently with four modules. And if you only use the PC for light office work, even a single 4GB or a single 8GB module would do. But in a PC where performance matters, for example for gaming, getting a single RAM module to upgrade when you have two existing modules is not good at all. The third module will be addressed in single-channel mode, while simultaneously ruining the memory system's electrical properties and making everything work at whatever the slowest module's specification is.

Note: When upgrading the RAM, it's always good to check for BIOS updates, they often improve compatibility with newer RAM modules (even if it's not explicitly mentioned in the changelog).


4) Today's sweet spot of DDR4-3600 with the latest CPUs

On AMD, DDR4-3600 has been the sweet spot for quite a while. But now, Intel introduced new memory controllers in their 11th gen and 12th gen CPUs which also require a divider above a certain RAM frequency. Only up to DDR4-3600 (but that pretty much guaranteed), the RAM and the CPU's memory controller (IMC) run at the same frequency (Intel calls this "Gear1 mode"). Somewhere above that RAM frequency, depending on the IMC's capabilities, the IMC has to resort to Gear2 mode, which introduces a divider for it and makes it run at half the RAM frequency. This costs a lot of performance.

An example on Intel Z590 with a kit of DDR4-3200: The IMC doesn't require a divider and can comfortably run in 1:1 mode (Gear1), which has the best performance.

BIOS OC.png


The Gear2 mode that becomes necessary at high RAM frequencies has a substantial performance penalty, because the latencies increase (everything takes a little longer). This basically leads to the same situation that we already know from AMD: RAM frequencies that are considerably above DDR4-3600 are almost useless, because of the divider being introduced for the IMC (memory controller). The performance loss with a divider is just too significant.

For the RAM performance to be on the same level again as DDR4-3600 without a divider (Gear1 mode on Intel), it requires something like DDR4-4400 (!) with the divider in place (Gear2 mode).

Looking at the high prices for DDR4-4400 kits or what it takes to overclock a normal kit of RAM to that, it's not practical. So with Rocket Lake (Core i-11000) and Alder Lake (Core i-12000) CPUs, and of course recent AMD CPUs, the "sweet spot" is usually at DDR4-3600. This frequency is known to not require a divider for the memory controller and thus gives the best performance and bang-for-buck.

Some more recent AMD CPUs, as well as 12th gen Intel "Alder Lake" CPUs, can sometimes go a bit above DDR4-3600 without requiring a divider for the memory controller.
But DDR4-3600 almost always runs well in 1:1 mode and has a better price/performance than RAM with higher specs, so it's still the top pick.

Here's an example of an AMD system (X570 with Ryzen 3900X). The tool HWinfo64 can show those frequencies in the "Sensors" window.
DDR4-3866 is too much to run in 1:1 mode, so the divider for the memory controller is active and performance is worse. DDR4-3600 manages to run in 1:1 mode and the performance is better.

View attachment 150421

The best thing on both platforms nowadays is to run DDR4-3600 without a divider and with some nice low timings if possible. Something like DDR4-4000 will usually make the BIOS enable the divider for the memory controller and it will be slower overall than DDR4-3600, despite the higher RAM frequency. This is because the latencies are effectively increased when the memory controller has to work at a lower frequency. With a DDR4-4000 kit of RAM for example, i would enable XMP, but then manually set a DRAM frequency of DDR4-3600. This should make the BIOS remove the divider for the memory controller and the performance will immediately be better.

Here's a page from an MSI presentation about 11th gen Rocket Lake CPUs, showing the increased latencies when the divider comes into play:
View attachment 158526

And here's from an AMD presentation about the Ryzen 3000-series, showing similarly increased latencies once the divider is active:
View attachment 159007


5) RAM stability testing

Memtest86 Free
from https://www.memtest86.com/
I use this as a basic stability test on a new system before i update the BIOS to the newest version (which is always one of the first things to do, as the factory BIOS will already be quite outdated).
Also, since it runs from a USB stick/drive, i use it as a first check before booting Windows, when something has significantly changed with the RAM or its settings.
One or two passes of this give me a good idea if the system is generally stable enough to either start installing Windows or boot it.
It's a good first test if you are completely unsure about stability, as well as a good "finisher" if you want to be extra sure that everything is ok with your memory system after doing other testing.
The main advantage is that it runs from USB. The main disadvantage is that RAM tests in Windows are more thorough in catching errors.
Launch the included ImageUSB program to prepare a USB drive with it, then boot from that drive (press F11 during POST for the boot menu).
Some people may also know "Memtest86+", a fork of Memtest86 which was better for a while. But by now, the regular Memtest86 is more current and the one to use.


Once in Windows, a quick way for detecting more obvious RAM instabilities is TestMem5 or TM5 for short: https://testmem.tz.ru/tm5.rar
Used with this config file (only the MT.cfg file is needed, it goes into the TM5\bin folder). TM5 delivers a good and not too time-consuming indication of RAM stability.
Run as admin. A full run of three passes with the config file from the link should take 15-40 minutes (the precise time it takes depends on the amount of RAM and the RAM/CPU performance).

Example of unstable RAM, with TM5 detecting six errors during testing:

1648934955114.png


Those weird letters on the bottom left are shown because that font doesn't contain the proper Cyrillic letters (but even if it did, most people couldn't read it).
Those lines say at which test (0, 8, 2, 2, 0...) the errors were found. But just make sure there are no errors being detected. If the RAM passes all tests, TM5 will say "of errors is not detected" at the end.


One of the best tools to thoroughly test RAM stability is from Google, and it's called GSAT (Google stressapptest). It has been specifically developed by Google to detect memory errors, because they use ordinary PCs instead of specialized servers for a lot of things. The only downside, it takes a bit of time to set up. To run GSAT, you first have to enable the "Windows Subsystem for Linux":



After the necessary reboot, open the Microsoft Store app and install "Ubuntu", then run Ubuntu from the start menu.
It will ask for a username and password, they are not important, just enter a short password that you remember, you need to enter it for the update commands.
Then run the following commands one after the other (copy each line, then right-click into the Ubuntu window to paste it, then press enter):

sudo apt-get update
sudo apt full-upgrade -y
sudo apt-get install stressapptest

Then you can start GSAT with the command:
stressapptest -W -M 12000 -s 3600

This example tests 12 GB of RAM (in case of 16 GB total, because you need to leave some for Windows), for 3600 seconds (one hour). You can also enter -s 7200 for two hours.
If you have more RAM, always leave 4 GB for Windows, so with 32 GB, you would use "-M 28000".
GSAT looks unspectacular, just some text scrolling through, but don't let that fool you, that tool is pretty stressful on your RAM (as it should be).
At the end, it has to say Status: PASS, and there should be no so-called "hardware incidents". Otherwise it's not stable.


然后,HCI记忆测试相当不错。有一个有用的工具,叫做MemTestHelper:https://github.com/integralfx/MemTestHelper/releases/tag/v2.2.0
它需要Memtest 6。4,可在此处下载:https:// www.3dfxzone.it/programs/?objid = 18508
(因为在最新的Memtest 7中。0,他们做了一个改变,这样MemTestHelper不再工作,你应该被迫购买Memtest Pro)。

将两个工具放在同一个文件夹中。启动MemTestHelper,使用16 GB RAM,您可以测试高达12000 MB (其余用于Windows)。
让它运行,直到通过400%。这是一个很好的指标,表明你的内存是稳定的。如果你真的想确定,让它运行到800%。

memtest_1.png


在严重的RAM超频中,另一个受欢迎的工具是卡胡https://www.karhusoftware.com/ramtest/
但是它要花10英镑€注册,所以我会使用其他免费程序 (除非你有RAM OC)。


稳定性测试也对内存控制器提出了很大挑战,因此对于完成与RAM相关的测试绝对有用:
Linpack Xtremehttps://www.techpowerup.com/download/linpack-xtreme/

运行Linpack,选择2 (压力测试),5 (10 GB),至少设置10次/次,按Y使用所有线程,然后让它执行它的操作。
这是检测不稳定的最佳工具之一,但警告说,这也会在CPU中产生大量热量。所以我会观察温度HWinfo64传感器。
最后必须说 “所有检查都通过了”,并且两个 “剩余” 数字列必须在整个试验 (运行) 中相同。
以下是三个试验的快速示例,显示了匹配的数字列:

View attachment 157277

“匹配数字列” 是指 “残差” 以下的所有数字必须匹配,“残差 (标准)” 以下的所有数字必须匹配。
因此,垂直匹配 (所有彼此下方的数字都是相同的)。只有两个数字水平方向不同。
我在图片中标记了两列。在每个绿色框中,数字必须相同,如图所示。“残差” 以下3倍,低于 “残差 (标准)” 下3倍。
通常你会进行10多次试验,而不仅仅是3次。每当有不同的数字时,尽管它显示 “通过”,内存系统并不是真正稳定的。

这是一个不同数字的例子,因此是一个不稳定的存储系统 (标记了不同数字的试验/运行):
View attachment 157285
仅检查 “残差” 和 “残差 (标准)”。“GFlops” 是一种性能指标,它是快速判断某些RAM调整是否有好处的好方法。当然,这个数字可能不同。


最后,一个关于内存和热量的重要注意事项: 较高的环境温度不利于RAM稳定性。在RAM特定压力测试中,RAM可能非常稳定。但是根据显卡 (功耗和冷却设计),一旦显卡在游戏过程中将热量排放到非常靠近RAM插槽的外壳中,可能会导致与RAM相关的崩溃,因为RAM会发热很多并失去稳定性。一种模拟方法是在进行压力测试时使用吹风机并将其松散地指向RAM。这正是许多显卡在负载下所做的事情,因为很多时候,冷却器的散热片都朝向主板和侧面板,所以热量从卡片的侧面散发出来,内存就在卡片的正上方。由此得出的结论是: 如果你的RAM在RAM压力测试中是好的,但是你在具有相同 (侵略性) RAM设置的游戏中崩溃了,你需要把它们松开一点,为这些情况增加一些净空。遵循RAM不稳定的三个原则: 放松计时和/或降低频率和/或提高电压。


深海潜水更多RAM中:
[URL unfurl = “true”] https:// forum-en。微星。com/索引。php?线程/让超频-一些-hynix-cjr。345491/# post-1999058[/URL]
[URL unfurl = “true”] https:// forum-en。微星。com/索引。php?线程/ddr-培训。345841/# post-2000832[/URL]
它可能很快会变得有点复杂,但是如果有任何问题,尽管问。
Since some people run into problems with four RAM modules on modern MSI mainboards, i wanted to explain the reasons behind that, and why two modules are often superior.
The main reason lies in the way the memory slots are connected to the memory controller, which is inside the CPU. So the first explanation is about:


1) RAM slot layout

All regular mainboards and desktop CPU models have a dual-channel memory system. Since a lot of boards offer four RAM slots, a pair of two slots have to each form a RAM channel.
So the four RAM slots are not individually addressed, but in pairs, as two channels. The different ways to connect the RAM slot pairs on the board are either "Daisy chain" or "T-Topology".
This RAM slot layout decision - the way the slots are connected - has a big influence on how many modules (two or four) the board works best with.

Here is a slide from an MSI presentation, showing that almost all of today's boards have a "daisy chain" memory slot layout. This layout heavily prefers two-module-operation.
The presentation is a bit older, but it's safe to say that the the vast majority of recent mainboards (B550, Z590, Z690 etc...) also have a daisy chain layout, and it's confirmed in several reviews.
Especially MSI are known to use this layout on almost all their modern boards. For other mainboard makers, it depends on the board model, but they will also tend to prefer this layout.

Daisy Chain.jpg


Daisy chain means that the slot pairs are connected one after the other, and therefore optimized for two modules total. The right slot of each channel is the end point.
Using two RAM modules, they are to be inserted into slot 2 and 4 counted from the left as per the mainboard manual. Meaning, into the second slot of each channel and thus the end point.
The reason is, this puts them at the very end of the PCB traces coming from the CPU, which is important for the electrical properties.
PCB (printed circuit board) traces are the thin signal lines that are visible on the mainboard, especially between the CPU and the RAM slots.

View attachment 149843

Why is this important? The PCB traces, going from the memory controller contacts of the CPU, to each contact of the RAM slots, are optimized to result in exactly the same distance between all those points. They are essentially "zig-zagging" across the board for an electrically ideal layout, making a few extra turns if a direct line would lead to an uneven distance.

This is done so that, with two modules, a) each RAM module is at the very end of the electrical traces coming from the CPU's memory controller, and b) each module has exactly the same distance to the memory controller across all contacts. We are dealing with nanosecond-exact timings, so all this matters.

On a mainboard with a daisy-chain RAM slot layout, this optimization is done with only two modules in mind, which are in slot 2 and 4 (on the board, those slots are called A2 and B2).
This is the configuration that most buyers would use, and it also results in the best overclocking potential. This way, the mainboard makers can boast with higher RAM overclocking frequencies when advertising the board, and the majority of buyers will have the ideal solution with two RAM modules.

Note: Never populate slots 1 and 3 first. When putting the modules into slot 1 and 3, the empty slots 2 and 4 would be like having some loose wires hanging from the end of each RAM contact, creating unwanted signal reflections and so on. So with two modules, they need to go into the second slot (slot 2+4, or A2 and B2) of each memory channel, to not have "loose ends" after the RAM module.

View attachment 155049


Now the interesting question. What happens when we populate all four slots on a mainboard with a daisy-chain slot layout? Well, the module in the second and fourth slot become "daisy-chained" after the modules in the first and third slot. This completely worsens the electrical properties of the whole memory system.

With four modules, there are now two modules per channel, and the two pairs of a channel don't have the same distance from the memory controller anymore. That's because the PCB traces go to the first slot, and then over to the second slot. This daisy-chaining - with the signal lines going to the first and then to the second module of a memory channel - introduces a lot of unwanted electrical handicaps when using four modules. The signal quality worsens considerably in this case.

Only with a "T-Topology" slot layout, the PCB traces have exactly the same length across all four slots, which would provide much better properties for four-module operation. But mainboards with T-Topology have gone a bit out of fashion, since most people use just two modules. Plus the memory OC numbers look much better with a daisy chain layout and two modules. So if the mainboard makers were to use T-topology on a board, they couldn't advertise with such high overclocking numbers, and people would think the board is worse (and it actually would be, for only two modules).

View attachment 156014
Example of an ASUS board with the rare T-Topology layout, advertising the fact that it works better with four modules compared to the much more common boards using the daisy-chain layout.


2) Single-rank vs. dual-rank

Another consideration is single-rank vs. dual-rank modules. This is about how a RAM module is organized, meaning, how the individual memory chips on the module are addressed. To put it simply, most (if not all) 8 GB modules are single-rank nowadays, as well as some 16 GB modules. A single-rank module puts less stress on the memory system. There's also a bunch of 16 GB modules that are dual-rank, and all bigger modules are always dual-rank. Dual-rank is slightly faster performance-wise (up to 4%), but also loads the memory controller more. One dual-rank module puts almost as much stress on the memory system as two single-rank modules!

What is the memory system? It consists of the CPU's integrated memory controller (IMC), the mainboard and its BIOS, and the RAM itself.

Every modern mainboard will be the happiest with two single-rank modules (for dual-channel operation), because this causes the least stress on the memory system, and is electrically the most ideal, considering that the memory slots are connected as "daisy chain". This fact is reflected in the maximum DDR frequencies that the mainboards are advertised with.

Here is an example from the highest MSI DDR4 board model using newest Intel Z690 chipset, which should therefore be quite advanced (specs of MPG Z690 EDGE WIFI DDR4, under "Detail"):
Max. overclocking frequency:
1DPC 1R Max speed up to 5200+ MHz
1DPC 2R Max speed up to 4800+ MHz
2DPC 1R Max speed up to 4400+ MHz
2DPC 2R Max speed up to 4000+ MHz

"DPC" means DIMM (=module) per channel, 1R means single-rank, 2R means dual-rank.

With 1DPC 1R = two single-rank modules (so, 2x 8 GB or 2x 16 GB single-rank), the highest frequencies can be reached.
With 1DPC 2R = two dual-rank modules (like 2x 16 GB dual-rank or 2x 32 GB), the maximum attainable frequency is lower, since the memory system is under more stress.
With 2DPC 1R = four single-rank modules (4x 8 GB or 4x 16 GB single-rank), the maximum frequency drops again, because four modules are even more challenging than two dual-rank modules.
And 2DPC 2R = four dual-rank modules (like 4x 16 GB dual-rank or 4x 32 GB) combines the downsides of the highest possible load on the memory controller with the electrical handicap of using four slots on a daisy-chain-mainboard.

The last configuration can already be difficult to get stable at DDR4-3200 sometimes, let alone DDR4-3600. One could consider themselves lucky to get DDR4-3600 working with four dual-rank modules, maybe having to use more relaxed timings for example. The 16 GB and 32 GB modules also often don't have particularly tight XMP timings to begin with, like DDR4-3600 18-22-22-42.
By the way: The second timing (tRCD) is more telling and important than the first one (tCL) to determine the module quality, but most people only look at the first one, CAS Latency.

With the new DDR5 standard, this drop in attainable frequency is even more pronounced. This is from the specs of one of the top MSI Z690 boards (specs of MEG Z690 ACE, under "Detail"):
Max. overclocking frequency:
1DPC 1R Max speed up to 6666+ MHz
1DPC 2R Max speed up to 5600+ MHz
2DPC 1R Max speed up to 4000+ MHz
2DPC 2R Max speed up to 4000+ MHz

When going from two modules (first and second entry) to four modules, the attainable frequency drops drastically. With two single-rank modules, DDR5-6000 and above is possible according to MSI. With two dual-rank modules, that goes down a little already. But with four modules, the memory system is under a lot more stress, and MSI are quite open about the result. This seems to be a limitation of the DDR5 memory system, which relies even more on a very clean signal quality. Using four DDR5 modules on a board with a daisy-chain layout clearly is not good in that regard.
This deterioration with four DDR5 modules is so drastic that the conclusion could be: DDR5 motherboards should come with only 2 dimm slots as standard (Youtube)


Generally, in case of RAM problems, no matter the technology, there are three possibilities, which can also be used in combination:
1) Raise voltage
2) Lower frequency
3) Loosen timings

But in some cases, buying different modules is actually the best solution.


3) Amount of RAM

32 or 64 GB RAM can be justified for large video editing projects, rendering, heavy photoshop use and such cases. But if it's gaming for example, modern games very rarely use more than 16 GB RAM. There are just the first games coming out, like Flight Simulator 2020, who will use a little more than 16 GB RAM, but still basically run at the same speed. So, 32 GB would definitely be a generous amount, but it can be befitting for a high-end gaming system. However, 64 GB amounts to a waste of money for gaming, no matter what. Before any game will ever touch more than 32 GB, the whole PC will be long outdated, because it could take years. Why are games satisfied with 16 GB nowadays? Consoles. A lot of games are developed with the lucrative console market in mind, and even the PlayStation 5 only has 16 GB of RAM. So games are designed from the ground up not to need more RAM, which then also applies to the PC versions of those games.

Recommendations for use on modern consumer mainboards:
8 GB RAM: Use 2x 4 GB, or even 1x 8 GB if RAM performance isn't critical anyway (entry-level systems, office work etc.)
16 GB RAM: Use 2x 8 GB (good mid-range / gaming systems, most of the time you won't need more)
32 GB RAM: Use 2x 16 GB (high-end gaming systems - when all other bottlenecks are removed - and uses beyond gaming, such as video processing etc.)
64 GB RAM: Use 2x 32 GB (uses beyond gaming), and consider using a professional quad-channel-memory platform like Threadripper with 4x 16 GB.
128 GB RAM total or higher: Use 4x 32 GB, but now you are maximally stressing the memory system in most configurations, so seriously consider using a professional quad-channel-memory platform like Threadripper with 4x 32 GB or more.

I usually recommend DDR4-3600 frequency with AMD and Intel Rocket Lake 11th gen / Alder Lake 12th gen (see chapter 4).


3b) How to increase the RAM size when you have 2x 4 GB or 2x 8 GB RAM?

First choice: Replace the 2x 4 GB with 2x 8 GB, or the 2x 8 GB with 2x16 GB. The new RAM should be a kit of matched modules. This will ensure the best performance and the least problems, because there's only two modules again in the end.

Second choice: Add a kit of two matching modules to your two existing modules. But you might not be able to get the same modules again. Even if they are the same model, something internally might have changed. Or you might toy with the idea of adding completely different modules (for example, adding 2x 8 GB to your existing 2x 4 GB). This can all cause problems. The least problems can be expected when you add two modules that are identical to your old ones. But then there's still this: You are now stressing the memory system more with four modules instead of two, so the attainable RAM frequency might drop a little. Also, it's electrically worse on a mainboard with daisy-chain layout, as explained under 1).

Lastly, adding just one more module (to have three modules total) is by far the worst choice for several reasons. Every desktop platform has a dual-channel memory setup. This means it works best with two modules, and it can work decently with four modules. And if you only use the PC for light office work, even a single 4GB or a single 8GB module would do. But in a PC where performance matters, for example for gaming, getting a single RAM module to upgrade when you have two existing modules is not good at all. The third module will be addressed in single-channel mode, while simultaneously ruining the memory system's electrical properties and making everything work at whatever the slowest module's specification is.

Note: When upgrading the RAM, it's always good to check for BIOS updates, they often improve compatibility with newer RAM modules (even if it's not explicitly mentioned in the changelog).


4) Today's sweet spot of DDR4-3600 with the latest CPUs

On AMD, DDR4-3600 has been the sweet spot for quite a while. But now, Intel introduced new memory controllers in their 11th gen and 12th gen CPUs which also require a divider above a certain RAM frequency. Only up to DDR4-3600 (but that pretty much guaranteed), the RAM and the CPU's memory controller (IMC) run at the same frequency (Intel calls this "Gear1 mode"). Somewhere above that RAM frequency, depending on the IMC's capabilities, the IMC has to resort to Gear2 mode, which introduces a divider for it and makes it run at half the RAM frequency. This costs a lot of performance.

An example on Intel Z590 with a kit of DDR4-3200: The IMC doesn't require a divider and can comfortably run in 1:1 mode (Gear1), which has the best performance.

BIOS OC.png


The Gear2 mode that becomes necessary at high RAM frequencies has a substantial performance penalty, because the latencies increase (everything takes a little longer). This basically leads to the same situation that we already know from AMD: RAM frequencies that are considerably above DDR4-3600 are almost useless, because of the divider being introduced for the IMC (memory controller). The performance loss with a divider is just too significant.

For the RAM performance to be on the same level again as DDR4-3600 without a divider (Gear1 mode on Intel), it requires something like DDR4-4400 (!) with the divider in place (Gear2 mode).

Looking at the high prices for DDR4-4400 kits or what it takes to overclock a normal kit of RAM to that, it's not practical. So with Rocket Lake (Core i-11000) and Alder Lake (Core i-12000) CPUs, and of course recent AMD CPUs, the "sweet spot" is usually at DDR4-3600. This frequency is known to not require a divider for the memory controller and thus gives the best performance and bang-for-buck.

Some more recent AMD CPUs, as well as 12th gen Intel "Alder Lake" CPUs, can sometimes go a bit above DDR4-3600 without requiring a divider for the memory controller.
But DDR4-3600 almost always runs well in 1:1 mode and has a better price/performance than RAM with higher specs, so it's still the top pick.

Here's an example of an AMD system (X570 with Ryzen 3900X). The tool HWinfo64 can show those frequencies in the "Sensors" window.
DDR4-3866 is too much to run in 1:1 mode, so the divider for the memory controller is active and performance is worse. DDR4-3600 manages to run in 1:1 mode and the performance is better.

View attachment 150421

The best thing on both platforms nowadays is to run DDR4-3600 without a divider and with some nice low timings if possible. Something like DDR4-4000 will usually make the BIOS enable the divider for the memory controller and it will be slower overall than DDR4-3600, despite the higher RAM frequency. This is because the latencies are effectively increased when the memory controller has to work at a lower frequency. With a DDR4-4000 kit of RAM for example, i would enable XMP, but then manually set a DRAM frequency of DDR4-3600. This should make the BIOS remove the divider for the memory controller and the performance will immediately be better.

Here's a page from an MSI presentation about 11th gen Rocket Lake CPUs, showing the increased latencies when the divider comes into play:
View attachment 158526

And here's from an AMD presentation about the Ryzen 3000-series, showing similarly increased latencies once the divider is active:
View attachment 159007


5) RAM stability testing

Memtest86 Free
from https://www.memtest86.com/
I use this as a basic stability test on a new system before i update the BIOS to the newest version (which is always one of the first things to do, as the factory BIOS will already be quite outdated).
Also, since it runs from a USB stick/drive, i use it as a first check before booting Windows, when something has significantly changed with the RAM or its settings.
One or two passes of this give me a good idea if the system is generally stable enough to either start installing Windows or boot it.
It's a good first test if you are completely unsure about stability, as well as a good "finisher" if you want to be extra sure that everything is ok with your memory system after doing other testing.
The main advantage is that it runs from USB. The main disadvantage is that RAM tests in Windows are more thorough in catching errors.
Launch the included ImageUSB program to prepare a USB drive with it, then boot from that drive (press F11 during POST for the boot menu).
Some people may also know "Memtest86+", a fork of Memtest86 which was better for a while. But by now, the regular Memtest86 is more current and the one to use.


Once in Windows, a quick way for detecting more obvious RAM instabilities is TestMem5 or TM5 for short: https://testmem.tz.ru/tm5.rar
Used with this config file (only the MT.cfg file is needed, it goes into the TM5\bin folder). TM5 delivers a good and not too time-consuming indication of RAM stability.
Run as admin. A full run of three passes with the config file from the link should take 15-40 minutes (the precise time it takes depends on the amount of RAM and the RAM/CPU performance).

Example of unstable RAM, with TM5 detecting six errors during testing:

1648934955114.png


Those weird letters on the bottom left are shown because that font doesn't contain the proper Cyrillic letters (but even if it did, most people couldn't read it).
Those lines say at which test (0, 8, 2, 2, 0...) the errors were found. But just make sure there are no errors being detected. If the RAM passes all tests, TM5 will say "of errors is not detected" at the end.


One of the best tools to thoroughly test RAM stability is from Google, and it's called GSAT (Google stressapptest). It has been specifically developed by Google to detect memory errors, because they use ordinary PCs instead of specialized servers for a lot of things. The only downside, it takes a bit of time to set up. To run GSAT, you first have to enable the "Windows Subsystem for Linux":



After the necessary reboot, open the Microsoft Store app and install "Ubuntu", then run Ubuntu from the start menu.
It will ask for a username and password, they are not important, just enter a short password that you remember, you need to enter it for the update commands.
Then run the following commands one after the other (copy each line, then right-click into the Ubuntu window to paste it, then press enter):

sudo apt-get update
sudo apt full-upgrade -y
sudo apt-get install stressapptest

Then you can start GSAT with the command:
stressapptest -W -M 12000 -s 3600

This example tests 12 GB of RAM (in case of 16 GB total, because you need to leave some for Windows), for 3600 seconds (one hour). You can also enter -s 7200 for two hours.
If you have more RAM, always leave 4 GB for Windows, so with 32 GB, you would use "-M 28000".
GSAT looks unspectacular, just some text scrolling through, but don't let that fool you, that tool is pretty stressful on your RAM (as it should be).
At the end, it has to say Status: PASS, and there should be no so-called "hardware incidents". Otherwise it's not stable.


Then, HCI Memtest is quite good. There is a useful tool for it, called MemTestHelper: https://github.com/integralfx/MemTestHelper/releases/tag/v2.2.0
It requires Memtest 6.4, which can be downloaded here: https://www.3dfxzone.it/programs/?objid=18508
(Because in the newest Memtest 7.0, they made a change so that MemTestHelper doesn't work anymore and you should be forced to buy Memtest Pro).

Put both tools in the same folder. Start MemTestHelper, and with 16 GB RAM, you can test up to 12000 MB (the rest is for Windows).
Let it run until 400% are passed. That's a good indicator that your RAM is stable. If you want to make really sure, let it run to 800%.

memtest_1.png


Another popular tool among serious RAM overclockers is Karhu from https://www.karhusoftware.com/ramtest/
But it costs 10€ to register, so i would just use the other free programs (unless RAM OC is your hobby).


A stability test which also challenges the memory controller a lot, and therefore definitely useful to round out the RAM-related testing:
Linpack Xtreme from https://www.techpowerup.com/download/linpack-xtreme/

Run Linpack, select 2 (Stress test), 5 (10 GB), set 10 times/trials at least, press Y to use all threads, and let it do its thing.
It's one of the best tools to detect instability, but warning, this also generates a lot of heat in the CPU. So i would watch the temperatures using HWinfo64 Sensors.
It has to say "All checks passed" at the end, and the two "residual" number columns have to be identical across the trials (runs).
Here's a quick example of three trials, showing matching number columns:

View attachment 157277

By "matching number columns", i mean that all the numbers below "Residual" have to match, and all the numbers below "Residual(norm)" have to match.
So, matching vertically (all the numbers below each other being identical). Only the two numbers horizontally to each other will be different.
I marked the two columns in the picture. In each green box, the numbers have to be the same, like in this picture. 3x the same below "Residual" and 3x the same below "Residual(norm)".
Normally you would run 10+ trials and not just three. And whenever there are a different numbers in there, then despite it showing "pass", the memory system is not truly stable.

This is an example of different numbers and therefore an unstable memory system (the trials/runs with the different numbers are marked):
View attachment 157285
Only check "Residual" and "Residual(norm)". "GFlops" is a performance metric, it's a good way to quickly judge if some RAM tuning has benefits or not. This number can be different of course.


Finally, an important note about RAM and heat: Higher ambient temperatures are not good for RAM stability. The RAM might be perfectly stable in a RAM-specific stress test. But depending on the graphics card (power consumption and cooling design), once the graphics card dumps its heat into the case very close to the RAM slots during gaming, there can be RAM-related crashes from the RAM heating up a lot and losing stability. A way to simulate this could be to use a hair dryer and point it loosely towards the RAM while doing stress-testing. This is precisely what many graphics cards are doing under load, because a lot of times, the fins of the cooler are oriented towards the mainboard and the side panel, so the heat comes out from the sides of the card, and the RAM sits right above that. The conclusion from this is: If your RAM is fine in RAM stress tests, but you have crashes in games with the same (aggressive) RAM settings, you need to loosen them up a bit to add some headroom for those circumstances. Go by the three principles of RAM instability: Loosen timings and/or lower frequency and/or raise voltage.


Deep-diving a bit more into RAM:
It can quickly become a bit complicated, but if there are any questions, feel free to ask.
Ask, which of the 3200C14 and 3600C16 performs well, how do I feel that the open XMP default 3200C14 is more smooth to play, but you say 3600 is the dessert frequency
[/报价]
 

7620772155602dd

New member
Joined
May 3, 2022
Messages
6
Please ask, 3200C14 and 3600C16 which performance is good, how do I feel that the XMP default 3200C14 to play the game is more smooth, but you said 3600 is the dessert frequency
 

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
DDR4-3200 CL14 works out to an absolute CAS latency of 8.75 ns.
DDR4-3600 CL16 works out to an absolute CAS latency of 8.89 ns.
This is because one clock cycle at 3600 MHz takes less time than one clock cycle at 3200 MHz.

So it's very similar, just 1.6% slower absolute CAS latency at DDR4-3600, you would never notice this alone.
But also: DDR4-3600 has a 12.5% higher bandwidth in read/write/copy transfers than DDR4-3200.

Now, how it can it be that DDR4-3200 CL14 might still "feel" smoother? Well, the RAM does not only consist of the transfer speed and the CAS latency.
There are dozens of other timings, and depending on the specific model of RAM, those other timings can be set quite differently and make more of a difference.

Even looking at four of the primary timings, there are remarkable differences in different kits/models of RAM, apart from the CAS Latency:
DDR4-3600 CL16-16-16-36
DDR4-3600 CL16-18-18-38
DDR4-3600 CL16-19-19-39
DDR4-3600 CL16-20-20-38

But a lot of secondary and tertiary timings can also be quite different between these kits. So then if you have a kit of DDR4-3200 CL14-14-14-34 and tighter other timings, it's possible that it feels slightly smoother in games than a kit of DDR4-3600 CL16-19-19-39 with looser other timings. Games like both: Higher transfer speeds and lower latencies. It depends not only on the CAS latency though, all the other timings are important too.
 

7620772155602dd

New member
Joined
May 3, 2022
Messages
6
DDR4-3200 CL14 works out to an absolute CAS latency of 8.75 ns.
DDR4-3600 CL16 works out to an absolute CAS latency of 8.89 ns.
This is because one clock cycle at 3600 MHz takes less time than one clock cycle at 3200 MHz.

So it's very similar, just 1.6% slower absolute CAS latency at DDR4-3600, you would never notice this alone.
But also: DDR4-3600 has a 12.5% higher bandwidth in read/write/copy transfers than DDR4-3200.

Now, how it can it be that DDR4-3200 CL14 might still "feel" smoother? Well, the RAM does not only consist of the transfer speed and the CAS latency.
There are dozens of other timings, and depending on the specific model of RAM, those other timings can be set quite differently and make more of a difference.

Even looking at four of the primary timings, there are remarkable differences in different kits/models of RAM, apart from the CAS Latency:
DDR4-3600 CL16-16-16-36
DDR4-3600 CL16-18-18-38
DDR4-3600 CL16-19-19-39
DDR4-3600 CL16-20-20-38

But a lot of secondary and tertiary timings can also be quite different between these kits. So then if you have a kit of DDR4-3200 CL14-14-14-34 and tighter other timings, it's possible that it feels slightly smoother in games than a kit of DDR4-3600 CL16-19-19-39 with looser other timings. Games like both: Higher transfer speeds and lower latencies. It depends not only on the CAS latency though, all the other timings are important too.
I'm talking about 3200C14 being 14-14-14-34, which is faster than the 3600 16-16-16 feeling
 

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
Again, the primary timings are comparable in latency, but you'd have to compare the secondary and tertiary timings, maybe there's an explanation in there.

Another possibility is that your board's BIOS has a better memory training result for your DDR4-3200 kit, meaning it is better optimized for that RAM and some miscellaneous RAM parameters are trained a bit tighter there, this goes even beyond the timings alone.

Also, i see that you are using a Z690 board (from your other posts). This platform is known to have a lot of trouble with the XMP profiles. One solution that was used there was to automatically apply Gear2 mode even for DDR4-3600 already, where it was not necessary yet on the previous platform of 11th gen / 500-series. So this is one thing to check as well: If DDR4-3200 runs in Gear1 mode (1:1 with memory controller) and DDR4-3600 runs in Gear2 mode (1:2 with the memory controller), then the explanation is clear: Gear2 mode increases all latencies by up to 30% or so!
 

The_King

New member
Joined
Feb 5, 2022
Messages
19
Hi guys,

Could someone please tell me if you can adjust CPU PLL or CPU +1.8V on the B450M Mortar Max?
I have gone through the BIOS and don't see the option to change this setting, even under advanced settings.

It is possible that lowering this to 1.7V can improve temps on RAM OC. I want to test if that is indeed true or not.

Thanks

My current RAM OC which is stable but RAM tends to get abit on the warm side with high ambient temps where I am.
3800 CL14 8GBX4 15V TM5 pass crop.jpg
 
Last edited:

Doc_Bucket

Member
Joined
May 12, 2022
Messages
59
I was talking more about the minimum for practical use which i have found to be the case. Because theoretically you can go lower than tRCD + tRP, so i should probably not call this a "formula" either, but it's just that in practice, i found it's not beneficial or gives problems when you go lower. As you said, you made similar observations. So i find this "formula" to be more correct than the other one.

For some timings there are minimums of course, meaning even if you set them lower, the memory controller will ignore your setting and just use the minimum. And a lot of timings are interconnected, so the minimum will be dynamic, i.e. depending on some other timing, or you may actually have to modify the other timing instead to get the initial timing to the value you want. The BIOS may regard those interconnections, which is good.
I've been going through Intel 10th gen Datasheet today and accidentally discovered 28T seems to be a hard lower limit for tRAS, regardless of other settings. So that was right, when I tried to set it lower, the IMC ignored my settings.

What I haven't grasped yet is the limitation of turnaround timings, in particular tWRRD. It is definitely in linear correlation to tWTR, but much higher; do you have an idea what makes part of that "turnaround cycle" and why these two must be differentiated?
 

citay

Pro
SERGEANT
Joined
Oct 12, 2016
Messages
10,815
On MSI, the tertiary tWRRD_sg and _dg timings are determined automatically during memory training, going off of the secondary tWTR_L and tWTR_S timings.
In the ASUS BIOS for example, it's the other way around: You set tWRRD_sg and _dg to get your desired tWTR_L and tWTR_S.
There are several ways how they are related, but on MSI you'd just set tWTR and tWTR_L, let's say 3 and 9, and let it determine the rest by itself, which should end up on _dg 23 or 24 and _sg 29.
 

Doc_Bucket

Member
Joined
May 12, 2022
Messages
59
On MSI, the tertiary tWRRD_sg and _dg timings are determined automatically during memory training, going off of the secondary tWTR_L and tWTR_S timings.
In the ASUS BIOS for example, it's the other way around: You set tWRRD_sg and _dg to get your desired tWTR_L and tWTR_S.
There are several ways how they are related, but on MSI you'd just set tWTR and tWTR_L, let's say 3 and 9, and let it determine the rest by itself, which should end up on _dg 23 or 24 and _sg 29.
Yes, this is actually what was happening. I started with tWTR_S and _L, always a little puzzled by tWRRD_ dg/_sg going up and down somehow by itself.

I was wondering because according to that Intel Datasheet, their lower limits are also 4, the same as tRDWR_dg/_sg, which will let me set them at 9. I've read elsewhere one of these two (i.e. either tWRRD or tRDWR) must be set significantly higher than the other, so I was not looking for lower limits in both cases. Yet, the question arose, could I go lower than AUTO.

Thank your for the assurance.
 
Joined
Jul 26, 2022
Messages
7
Think it's safe to say I'm pretty much invested in this whole RAM overclocking hobby now. So, I came here to ask a genuine question: I've got the 4400 Patriot Viper kit running at 3600 14-14-14-29 Gear 1 at 1.45v. It's B-die so I figured that's it's not too high a voltage to daily. It's stable but I would like to go further - I did boot with 3700 14-14-14-32 but BSOD while gaming. Does anyone have any recommendations as to what I should focus on with this kit? It runs the XMP 4400 profile, but I haven't gone further. I feel as though I'm only halfway done with finding the limit to this kit.
 
Top