LabRat 891Good luck with that, AMD.

Literally, the ambient temperature, weather, and (especially) EMI and trace layout, directly impact Stability and RAM timings.

The signal travel delay from CPU to memory chips and back is around 1.2 ns (at 20 cm/ns and 12 cm distance). That's a minor source of latency. The major source is what's happening in the DRAM chip itself. For each read or write operation, electrons move from capacitor to capacitor several times through transistors and wires which have some resistance. That is slow, the IMC can't do anything but wait, and latencies tell it how long. The total wait is the sum of all four primary latencies or something like that. Luckily, this operation is done on entire rows of 8 KB or so, so multiple reads within a row are much faster (tCL only), and sequential reads are much faster still.

LabRat 891That's the entire reason why "Memory Training" became standard in the first place; and, why a weeks-months Stable tweak/OC can suddenly go unstable.

Yes. This is link training, to be exact. The transmitting and receiving circuits (PHY) adapt to the electrical characteristics of the link (=wires) by trial and error until they achieve error-free transmission. It's similar in PCIe (at least since Gen3) and DDR (at least since DDR3) but PCIe stays at a lower speed if a higher one doesn't work while DDR stays at one pre-set frequency.
The result of training are probably some optimised phase shifts or "latencies" in small fractions of a clock cycle, but they are inaccessible to us and not related to the timings we know.
Apart from temperature etc., the contact resistance in memory slots may change due to time/vibration/who knows what, and again error-free transmission collapses until new link training.

rv8000Send me a kit :). I promise I won’t make blanket generalizations about timings.

Perfect, I contact Origin Code on your behalf right away.

Don't forget to do some actual thermal testing while your at it. Also you'll have to return the kit when your done since it's a expensive sample. Other people need it after you :)

ir_cowPerfect, I contact Origin Code on your behalf right away.

Don't forget to do some actual thermal testing while your at it. Also you'll have to return the kit when your done since it's a expensive sample. Other people need it after you :)

Great, I await further details.

rv8000Send me a kit :). I promise I won’t make blanket generalizations about timings.

Yeah no, that'd be a waste of shipping costs.

ToothlessYeah no, that'd be a waste of shipping costs.

That’s cute.

I’ve spent countless hours tuning kits on AM5, offering timing advice here and on OCN, pages of written notes on voltage timings, speeds, termination values, 1:1, 2:1 testing, m-die vs b-die, 24+ validation testing on those overclocks and so on, but sure maxing trefi on AM5 with no direct cooling just immediately crashes kits…

Obviously ir is being facetious and wouldn’t actually have a kit sent out. Not much stock to place in claims that arent even investigated in memory reviews because they don’t have the time or want to spend the time to validate in depth timing customization.

rv8000That’s cute.

I’ve spent countless hours tuning kits on AM5, offering timing advice here and on OCN, pages of written notes on voltage timings, speeds, termination values, 1:1, 2:1 testing, m-die vs b-die, 24+ validation testing on those overclocks and so on, but sure maxing trefi on AM5 with no direct cooling just immediately crashes kits…

Obviously ir is being facetious and wouldn’t actually have a kit sent out. Not much stock to place in claims that arent even investigated in memory reviews because they don’t have the time or want to spend the time to validate in depth timing customization.

That's also cute.

Expecting a reviewer to spend countless hours on every single ram kit is asinine. There is zero real world reason for someone to spend that much time for the small crowd of memory tuners. Just because him or other reviews aren't catering to your wants doesn't mean the attitude is needed. Do your own reviews and see how worth the time is in the end.

rv8000Obviously ir is being facetious and wouldn’t actually have a kit sent out. Not much stock to place in claims that arent even investigated in memory reviews because they don’t have the time or want to spend the time to validate in depth timing customization.

This is true. I was being a bit overly sarcastic. I'm trying to get better at keeping in mind a keyboard warrior are the loudest because they care the most like yourself. I won't apologize though.

Anyways, the reviews I write focus on the average consumer, not the 400 people on OCN that go sub-ambient. The goal of the overclocking section in each review is easy to digest information and baby-steps into memory overclocking. The least amount of variables is the key here. Once you need a specific motherboard, a golden bin IMC and a unsafe voltages it is something that becomes unattainable for 99.9% of the readers. Also factoring DRAM IC characteristics, PCB type (KO or Brainpower), layer count and that also just adds to the list of specific set of circumstances for a stable memory overclock.

One example would be that during the G.SKIL 2026 CUP I submitted 10K freq for UDIMM. Used 7200 memory, which was completely unstable at that speed. Because it doesn't have to be stable for any of those entries, just have to last long enough to submit. If I were to say 10K is obtainable for all everyone with that 7200 kit, that would be a lie. If I provided all the settings and details, it would mostly end up being unless information too since it was platform, motherboard, CPU and voltage specific. In this scenario I spent a few days working on different submissions. That is time which could be spent better elsewhere and not for those 400 people who will never read a review in the first place. Instead I often observe they complain on forums that someone else isn't going enough in-dept.

To bring it back all the way to my original comment that high (65K( tREFi is a bad idea in a EXPO profile. It is because 99% of users just enable EXPO / XMP and play games. Those WHEA errors will cause a lot of problems. They are called silent errors for a reason. Maybe even further to the point Selaya comment sums the need to perform advanced clocking up perfectly; "x3d exists". I will always tell someone to buy a x3d before wasting time with AMD memory overclocking.

Hilux SSRGHow is this different than me doing trial and error on cas and timings manually in bios to reach stable low latency, other than paying a premium for the factory to do this?

You answered your own question. Not everybody likes sitting in front of the BIOS and fiddling with memory testing applications for hours or days just to find something that works. But on the other hand, those people (myself included) probably don't care about super tight timings, either, so there's that.

ir_cowThis is true. I was being a bit overly sarcastic. I'm trying to get better at keeping in mind a keyboard warrior are the loudest because they care the most like yourself. I won't apologize though.

Anyways, the reviews I write focus on the average consumer, not the 400 people on OCN that go sub-ambient. The goal of the overclocking section in each review is easy to digest information and baby-steps into memory overclocking. The least amount of variables is the key here. Once you need a specific motherboard, a golden bin IMC and a unsafe voltages it is something that becomes unattainable for 99.9% of the readers. Also factoring DRAM IC characteristics, PCB type (KO or Brainpower), layer count and that also just adds to the list of specific set of circumstances for a stable memory overclock.

One example would be that during the G.SKIL 2026 CUP I submitted 10K freq for UDIMM. Used 7200 memory, which was completely unstable at that speed. Because it doesn't have to be stable for any of those entries, just have to last long enough to submit. If I were to say 10K is obtainable for all everyone with that 7200 kit, that would be a lie. If I provided all the settings and details, it would mostly end up being unless information too since it was platform, motherboard, CPU and voltage specific. In this scenario I spent a few days working on different submissions. That is time which could be spent better elsewhere and not for those 400 people who will never read a review in the first place. Instead I often observe they complain on forums that someone else isn't going enough in-dept.

To bring it back all the way to my original comment that high (65K( tREFi is a bad idea in a EXPO profile. It is because 99% of users just enable EXPO / XMP and play games. Those WHEA errors will cause a lot of problems. They are called silent errors for a reason. Maybe even further to the point Selaya comment sums the need to perform advanced clocking up perfectly; "x3d exists". I will always tell someone to buy a x3d before wasting time with AMD memory overclocking.

And to reiterate, 65k is fine with passive cooling. TRFC and tREFI tuning are the easiest sources of latency improvements on the platform. Your reviews don’t cover it and you have otherwise not tested it or shown proof. While I don’t have any dual rank kits to test, every single kit of DDR5 I’ve had since my 7700x and 7900X3D have been run with no direct cooling at max trefi after 24 hour+ stability tests with no issues.

I’m not trying to devalue the work of the reviews, and have already stated it’s likely not within time constraints or reviewers have the ability to spend time checking extensive timing tuning, simply that you are wrong about tREFI.

I'll leave it here – on the topic of "heat" : )

rv8000And to reiterate, 65k is fine with passive cooling. TRFC and tREFI tuning are the easiest sources of latency improvements on the platform. Your reviews don’t cover it and you have otherwise not tested it or shown proof. While I don’t have any dual rank kits to test, every single kit of DDR5 I’ve had since my 7700x and 7900X3D have been run with no direct cooling at max trefi after 24 hour+ stability tests with no issues.

Well there is a correlation between one CCD vs 2x CCD. A single cannot fully saturated the memory leading to lower temps overall. This a curd preliminary test I did last month to narrow down what I should focus on for the next DRAM temp article.

rv8000I’m not trying to devalue the work of the reviews, and have already stated it’s likely not within time constraints or reviewers have the ability to spend time checking extensive timing tuning, simply that you are wrong about tREFI.

I will admit that I don't not go indepth per review on the timings I adjust. I figured people can do some research on their own if they cared enough. The extent is just saying what I adjusted and why is about it. Something I'll see if adding per article makes sense. More likely I'll just link to the bigger temp article I'm working on right now.

Anyways, saying 65K trefi will never be a issue is just simply wrong. Excluded external factors like GPU heat dumping all over the memory or case airflow, frequency and voltage independently contribute to dram temperature. So for example a 8000 1.45v kit will have different temps than 6000 1.25v. Add in the heatspeaders to the equation and you now have a lot of factors to consider. If you want to cherry pick examples, go for it. But lots of data for me and other people point to you cannot just universally say it's not issue.

ir_cowWell there is a correlation between one CCD vs 2x CCD. A single cannot fully saturated the memory leading to lower temps overall. This a curd preliminary test I did last month to narrow down what I should focus on for the next DRAM temp article.





I will admit that I don't not go indepth per review on the timings I adjust. I figured people can do some research on their own if they cared enough. The extent is just saying what I adjusted and why is about it. Something I'll see if adding per article makes sense. More likely I'll just link to the bigger temp article I'm working on right now.

Anyways, saying 65K trefi will never be a issue is just simply wrong. Excluded external factors like GPU heat dumping all over the memory or case airflow, frequency and voltage independently contribute to dram temperature. So for example a 8000 1.45v kit will have different temps than 6000 1.25v. Add in the heatspeaders to the equation and you now have a lot of factors to consider. If you want to cherry pick examples, go for it. But lots of data for me and other people point to you cannot just universally say it's not issue.

Why are you still beating around the bush? Show me a test with a kit set to 65k tREFI thats tested for stability on ambient - as this was the claim. No need to move the goal posts and start talking nonsense about single vs dual ccd chips (those were merely the cpus I had at the time I was tuning on AM5 first).

Here I even found an example I have on my phone…

rv8000Why are you still beating around the bush? Show me a test with a kit set to 65k tREFI thats tested for stability on ambient - as this was the claim.

Well since your asked :)

I figured the most obvious one is Origin Code which with 4 DIMMs errored out at 6000 EXPO. Even with the beefy heatspeaders, heat buildup in the center of 4 DIMMs overtook it. Without a fan directly blowing down on it there is no way 65K or any OC would work.

Now you might say that is edge case and only happens to 4 DIMMs. I will present to you XPG 8000 Neon kit. The plastic heatspeaders does nothing and at XMP it hits 73c. I actually didn't write a overclocking section specifically because even with a fan, it was barely holding on.

www.techpowerup.com/review/xpg-lancer-neon-rgb-ddr5-8000-32-gb-cl38/3.html

www.techpowerup.com/review/origin-code-vortex-ddr5-6000-192-gb-cl26/10.html

Those are two recent-ish reviews. But generally the reason I put a fan on the memory is because they will error out otherwise.

Now there is still research to be done on just tREFi applied and nothing else. But I think once again frequency and voltage will play a role in that outcome. Good thing to test and is already part of my upcoming article.

rv8000No need to move the goal posts and start talking nonsense about single vs dual ccd chips (those were merely the cpus I had at the time I was tuning on AM5 first).

No goal post moved. I was simply trying to figure out how you don't get errors. Lots of factors to consider, but I figured it's because of just using a 8core (1xCCD) instead of 16core for AMD

The fact that I and many others have the same end result, already disproves your stance that no memory can error out at 65K.

ir_cowplastic heatspeaders

Is that really a thing? :eek: Oh my god, enshittification is real, and they aren't even hiding it anymore!:fear:

ir_cowWell since your asked :)

I figured the most obvious one is Origin Code which with 4 DIMMs errored out at 6000 EXPO. Even with the beefy heatspeaders, heat buildup in the center of 4 DIMMs overtook it. Without a fan directly blowing down on it there is no way 65K or any OC would work.

Now you might say that is edge case and only happens to 4 DIMMs. I will present to you XPG 8000 Neon kit. The plastic heatspeaders does nothing and at XMP it hits 73c. I actually didn't write a overclocking section specifically because even with a fan, it was barely holding on.

www.techpowerup.com/review/xpg-lancer-neon-rgb-ddr5-8000-32-gb-cl38/3.html

www.techpowerup.com/review/origin-code-vortex-ddr5-6000-192-gb-cl26/10.html

Those are two recent-ish reviews. But generally the reason I put a fan on the memory is because they will error out otherwise.

Now there is still research to be done on just tREFi applied and nothing else. But I think once again frequency and voltage will play a role in that outcome. Good thing to test and is already part of my upcoming article.


No goal post moved. I was simply trying to figure out how you don't get errors. Lots of factors to consider, but I figured it's because of just using a 8core (1xCCD) instead of 16core for AMD

The fact that I and many others have the same end result, already disproves your stance that no memory can error out at 65K.

Except you’re not testing tREFI, you’re testing the stock kit, and essentially all youve shown is on whatever cpu/mobo combo you tested on + whatever primaries and secondaries are being set (bios randomly filling some timings) is not stable. Nor have you likely bothered to actually check if termination values or slightly higher vdd/vddq (or vddp for 2:1) is required to actually have a stable kit. This does not in any world mean tREFI at 65k is or will be the cause of instability. Merely that your test setups have not been stable when testing on AM5, which to be fair is different conversation.

I don’t think we know what these new profiles will include timing wise. Will it be all primaries, secondaries and tertiaries? Or piece meal with random infill from the mobo vendor bios.

TLDR:

An unstable kit from the get go is going to be unstable no matter what tREFI is set. Which factually has nothing to do max tREFI being perfectly stable at ambient when your kit/setup actually has proper settings.

rv8000An unstable kit from the get go is going to be unstable no matter what tREFI is set. Which factually has nothing to do max tREFI being perfectly stable at ambient when your kit/setup actually has proper settings.

What do you not understand here? Max tREFi is not stable if the temperature is to high. That is universally true for all memory.

TLDR: tREFi affects DRAM tempature. High enough temperature and the DRAM errors out. I showed two examples and you simply dont want to accept factual information.

rv8000Except you’re not testing tREFI, you’re testing the stock kit, and essentially all youve shown is on whatever cpu/mobo combo you tested on + whatever primaries and secondaries are being set (bios randomly filling some timings) is not stable. Nor have you likely bothered to actually check if termination values or slightly higher vdd/vddq (or vddp for 2:1) is required to actually have a stable kit. This does not in any world mean tREFI at 65k is or will be the cause of instability. Merely that your test setups have not been stable when testing on AM5, which to be fair is different conversation.

I don’t think we know what these new profiles will include timing wise. Will it be all primaries, secondaries and tertiaries? Or piece meal with random infill from the mobo vendor bios.

TLDR:

An unstable kit from the get go is going to be unstable no matter what tREFI is set. Which factually has nothing to do max tREFI being perfectly stable at ambient when your kit/setup actually has proper settings.

Run four dual ranks at max and see what happens. We'll wait since you want to argue.

ir_cowWhat do you not understand here? Max tREFi is not stable if the temperature is to high. That is universally true for all memory.

TLDR: tREFi affects DRAM tempature. High enough temperature and the DRAM errors out. I showed two examples and you simply dont want to accept factual information.

ir_cowI thought that was just EXPO 1.2 spec. If they really are planning on adding tREFi to its own profile, better start including a fan with every kit :)

In neither example did you show:

- value of tREFI (or full zen timings for that matter
- adjusting tREFI
- checking other timings or voltage to ensure stability

But somehow come to the conclusion tREFI is the only reason those kits crashed.

Meanwhile I provided a ~12 hour stability test including hwinfo spd read temps (which are significantly lower than what physical probes will provide) with ambient cooling and no direct fan on the kit, with zero errors.

You have a whole mess of variables not accounted for. The argument was that tREFI at 65k can in fact be run on ambient without errors. It wasnt here’s this kit with plastic heatspreaders, im not gonna show what tREFI is, im not gonna verify the kit is stable, and im not gonna run a sweep of tREFI values to see how its tied to stability.

@rv8000 you should go back and read my original statement. Maybe everything else for that matter because you are still not understanding.

In those two examples tREFi is not the reason they crashed. I picked those to illustrate that when memory ICs reach a certain temperature, they produce errors. So how could adding tREFi on top of that make it stable? It wouldn't. Which is a counter to one of your arguments that DRAM temps is never a problem and it's my OC timings that is the reason issue. Except that is stock XMP/EXPO. Nothing I did.

Which bring us to my original comment and point of this hand holding contest. If that XPg Neon kit is at the edge just with XMP, adding 5c+ because of tREFi will produce errors.

Round this all out. I said having 65K tREFi in the ULL profile would be a bad idea without a fan on the memory. It is because the amount of variables is immense. A blanket statement saying tREFi will never be a issue is just incorrect.

ir_cow@rv8000 you should go back and read my original statement. Maybe everything else for that matter because you are still not understanding.

In those two examples tREFi is not the reason they crashed. I picked those to illustrate that when memory ICs reach a certain temperature, they produce errors. So how could adding tREFi on top of that make it stable? It wouldn't. Which is a counter to one of your arguments that DRAM temps is never a problem and it's my OC timings that is the reason issue. Except that is stock XMP/EXPO. Nothing I did.

Which bring us to my original comment and point of this hand holding contest. If that XPg Neon kit is at the edge just with XMP, adding 5c+ because of tREFi will produce errors.

Round this all out. I said having 65K tREFi in the ULL profile would be a bad idea without a fan on the memory. It is because the amount of variables is immense. A blanket statement saying tREFi will never be a issue is just incorrect.

A blanket statement saying tREFI at max requires active cooling is factually incorrect. I can agree there are plenty of variables, but you have failed to provide an example in which tREFI was actually investigated after providing a stable baseline.

You’re trying to spout as opinion ad fact with no verified proof; you plugged some DDR5 kit into a board, slapped the expo button and attached some thermistors. Zero timing info, zero verification of stability (you probably had a tphy missmatch between dimms on the 8000 kit which means theyre going to randomly produce errors at different time intervals reboot to reboot no matter what you do; check vddp, adptinrv, or termination values). Do I really have to explain a good basis for scientific testing?

If vendors take the time to provide a robust profile with the proper bios values, 1:1 kits would likely have very few issues (IMC quality generally plays a bigger role with 2:1 kits, different beast). Refresh window is the lowest hanging fruit on the AM5 platform for latency improvements, of course that’s where they’re most likely to focus on for improvements.

rv8000A blanket statement saying tREFI at max requires active cooling is factually incorrect. I can agree there are plenty of variables, but you have failed to provide an example in which tREFI was actually investigated after providing a stable baseline.

You’re trying to spout as opinion ad fact with no verified proof; you plugged some DDR5 kit into a board, slapped the expo button and attached some thermistors. Zero timing info, zero verification of stability (you probably had a tphy missmatch between dimms on the 8000 kit which means theyre going to randomly produce errors at different time intervals reboot to reboot no matter what you do; check vddp, adptinrv, or termination values). Do I really have to explain a good basis for scientific testing?

If vendors take the time to provide a robust profile with the proper bios values, 1:1 kits would likely have very few issues (IMC quality generally plays a bigger role with 2:1 kits, different beast). Refresh window is the lowest hanging fruit on the AM5 platform for latency improvements, of course that’s where they’re most likely to focus on for improvements.

And a blanket statement would be you're literally ignoring what is said so "you can be right" throughout everything. Good job kid. You've proven you cannot understand the text being put in front of you.

rv8000A blanket statement saying tREFI at max requires active cooling is factually incorrect. I can agree there are plenty of variables, but you have failed to provide an example in which tREFI was actually investigated after providing a stable baseline.

I already proved you wrong on this and I can keep pointing back to the Neon kit. IF it is just barely stable at XMP how could it possibly be okay with max tREFi? So that is a example of memory that DOES need a fan for tREFi. Does this apply to all memory kits? No, but it does to some.

rv8000You’re trying to spout as opinion ad fact with no verified proof; you plugged some DDR5 kit into a board, slapped the expo button and attached some thermistors. Zero timing info, zero verification of stability (you probably had a tphy missmatch between dimms on the 8000 kit which means theyre going to randomly produce errors at different time intervals reboot to reboot no matter what you do; check vddp, adptinr Do I really have to explain a good basis for scientific testing?

The question is do I need to explain scientific testing to you ?

I'm m starting to think you never had read anything I've written. These insults fall flat and a bit funny to me you have run out of actual debatable tactics that this is all you have left. We are at a standstill here. What I presented is ignore. You keep asking for proof and objectively speaking I have done that.

Excluding external variables like GPU heat or case airflow, it is straightforwards. Frequency and voltage both contribute to dram ic temperature. Memory can have multiple ranks as well. Not all vendors are the same in heatspeaders quality or design.

So I'll say this again. Passive cooled memory can have a issue with max tREFi (AMD). I recommend a fan because of all the other variables I mentioned already.

The ideal setup would be single rank, low voltage memory to use ULL. Very low chance that will be a issue. On the other hand 6400 CL28 1.45V should have some active cooling for ULL or other sub-timining adjustments. Otherwise you run the risk of instability.

rv8000If vendors take the time to provide a robust profile with the proper bios values, 1:1 kits would likely have very few issues (IMC quality generally plays a bigger role with 2:1 kits, different beast). Refresh window is the lowest hanging fruit on the AM5 platform for latency improvements, of course that’s where they’re most likely to focus on for improvements.

ULL is a good idea overall. Probably a safer option is tREFi 32K. Get similar results, better temps. Less issues for consumers.

ir_cowI already proved you wrong on this and I can keep pointing back to the Neon kit. IF it is just barely stable at XMP how could it possibly be okay with max tREFi? So that is a example of memory that DOES need a fan for tREFi. Does this apply to all memory kits? No, but it does to some.


The question is do I need to explain scientific testing to you ?

I'm m starting to think you never had read anything I've written. These insults fall flat and a bit funny to me you have run out of actual debatable tactics that this is all you have left. We are at a standstill here. What I presented is ignore. You keep asking for proof and objectively speaking I have done that.

Excluding external variables like GPU heat or case airflow, it is straightforwards. Frequency and voltage both contribute to dram ic temperature. Memory can have multiple ranks as well. Not all vendors are the same in heatspeaders quality or design.

So I'll say this again. Passive cooled memory can have a issue with max tREFi (AMD). I recommend a fan because of all the other variables I mentioned already.

The ideal setup would be single rank, low voltage memory to use ULL. Very low chance that will be a issue. On the other hand 6400 CL28 1.45V should have some active cooling for ULL or other sub-timining adjustments. Otherwise you run the risk of instability.


ULL is a good idea overall. Probably a safer option is tREFi 32K. Get similar results, better temps. Less issues for consumers.

You didn’t prove anything wrong about the statement “max tREFI can be run at ambient”with the XPG kit review. You don’t list what software was used to check for errors and the only thing stated about temperature is that you ran a “preliminary thermal test”.

How do you prove a statement wrong when you’re not even testing or attempting to disprove what’s being discussed. You quite literally never test changing tREFI in that or the other review under ANY system condition or variables. Not to mention testing 2:1 kits on an AM5 platform is asking for issues from the get go.

Meanwhile I show you a stability test in which a kit is being run without any direct cooling in a closed system, at max tREFI AND running at significantly tighter timings and higher voltages than stock. I’ve run every single kit of DDR5 on AM5 in these conditions at max tREFI between the dozen or so G.skill, Corsair and T-Create kits i’ve owned and tested (primarily in looking for good kits to run in 2:1).

Take the L and move on, or at the very least get back to me when you ACTUALLY test whats being questioned here - can tREFI at 65k be run with no direct cooling.

@rv8000 you are not a very good listener....

I understand what your asking to "prove", so who's moving the goal post now? Let's admened your statement to say only 1:1 ratio, single rank and low voltage will be fine with max tREFi lol

ir_cow@rv8000 you are not a very good listener....

Nor you. How many times must it be explained that in neither review are you testing what happens with tREFI being changed, or for that matter testing tREFI.

Not to mention I’ve actually provided proof, that is conveniently being ignored. Where is your stability test with tREFI at 65k?