[Sammelthread] Intel DDR5 RAM OC Thread

Yes, should be okay, but pden 2 could mess the powerdown. Maybe it's not a good idea. It's a Buldzoid thing. 🙃
Its hard to believe this would be a recommendation
Maybe its just a reading mistake

RBL & WBL = 16
This is not a speed timing, this is a hardware design limitation.
DDR2 was RBL/WBL = 4
1705775282720.png
 
Wenn Du diese Anzeige nicht sehen willst, registriere Dich und/oder logge Dich ein.
Buildzoid use this in his YouTube videos. And many users just copy that.
I think its just something as simple as a "late night reading mistake".
RPDEN
RDPDEN
PRPDEN
all look nearly the same :giggle:
 
Zuletzt bearbeitet:
I think its just a late night reading mistake.
The value difference is too big.

You can not run neither CAS nor CWL at 2.
I know it's hard to believe, but you can see it in most of his videos.

In the meantime I started to play with the RTT's. The combination you suggested got error code 55. Probably due to the low voltages.
I got a tip from @CarSalesman last year, so tried it. 34-120-80-40-40
I was able to boot it, but got error 0 iin 20. minutes. So I upped the Dqs to 48, now I got this:
1705777568115.png

Need more tuning, but could be a good base.
 
Zuletzt bearbeitet:
Guten Abend!

Die Profis redet über extreme OC, ich hole jetzt andere Ergebnisse ;)
Ich habe eine ziemlich gute 13700k, heute eingebaut.
Ich denke, es schaut nicht so schlegt aus. Ja ok, nicht so kraftig wie 14900k, aber gehts.
Delided, IHS drauf, LM ist zwischen Chip und IHS.
WLP ist zwischen Block und CPU.
13700k ist allcore 53p 42e 45r, und 2 Kerne boosted 54x.
Ich habe die ganze CPU Einstellungen so gelassen, wie Default ist. Nur habe ich die Best Case Scenario benutzt, die ein sehr gute UV ist.
Alle andere sind Default. Es gibt kein noch V/F, und solche Dinge.
Vcore getested kurze SFT, wenn es passt, dann alles passt. (Erste Bild)
Es war mit 5600 JEDEC, und danach 8000MHz eingegeben.
25 Rund TM5 ist sofort gegangen, und nacher VST/VT3 habe gestarted.
Wow, es hat mich gewundert, 50 Mins is gelaufen.
Nochmal, noch kein OC gibt, alles ist Default, nun 8000 plus paar Werte eingegeben.
Temperaturen sind auch schön :)

Ich mache weiter dann, nie aufgeben ;)

Schöne Nacht u. WE an Alle!
Prost, zeb
Beitrag automatisch zusammengeführt:

Update:
Noch paar Werte eingegeben, und genau 50. Minutes Error. Immer VT3... grr, ich mag kein VT3 Test :fire:
 

Anhänge

  • Default-BIOS-13700k-SFT-VST-VT3.jpg
    Default-BIOS-13700k-SFT-VST-VT3.jpg
    694,6 KB · Aufrufe: 82
  • Default-BIOS-13700k-8000-C36-TM5-VST-VT3.jpg
    Default-BIOS-13700k-8000-C36-TM5-VST-VT3.jpg
    730,9 KB · Aufrufe: 86
  • 8000_neu.jpg
    8000_neu.jpg
    762,8 KB · Aufrufe: 84
Zuletzt bearbeitet:
Soviel wie ich weiss und was ich gesehen habe der erste Daily Screen auf einem Tachyon Z790 @ 8533 Mhz :coolblue:

Rams sind wassergekühlt mit einem Bykski RAM-Cooler

( Das war wirklich viel Zeit und Arbeit ).......werde das noch weiter verfestigen ....weil später kam leider noch ein Fehler :rolleyes: bei ~ 1:20 Std.


PS: Ich hatte ja mal vor einer Zeit gefragt was es überhaupt bringt auf Wasserkühlung für die RAMs zu bauen .

Aus jetziger Erfahrung kann ich sagen das das enorm Stabilität bringt. Das hätte ich früher machen sollen . Das hätte mir viel Zeit ersparen können.

Eventuell kommt noc heine ICEMAN RAM Cooler ...um das Delta noch weiter zu senken.

Mit dem Bykski habe ich ein Delta von ca. 7-8°c zur Wassertemp
 

Anhänge

  • TG 8533.png
    TG 8533.png
    558,7 KB · Aufrufe: 89
  • 20240120_012854.jpg
    20240120_012854.jpg
    949,4 KB · Aufrufe: 89
  • Capture001.png
    Capture001.png
    328,8 KB · Aufrufe: 87
Zuletzt bearbeitet:
80 WR is weaker

Then 80-80-60-40-48
For RD I need low (strong) value to even boot it. (48 is the highest which can boot) Is that means I already have high voltage on my dimms?
Should I lower the Voltage and increase the RD to 80?
Current state on 34-80-60-40-48. It's still running. RX-DFE on auto atm.
I was looking for the detailed comment from you about RTT's, but can't find it. Should be bookmarked somewhere I believe.

Let's see if lowered Dqs fix it:

1705812169716.png
 

Anhänge

  • 1705811372805.png
    1705811372805.png
    177,9 KB · Aufrufe: 51
Zuletzt bearbeitet:
Hi @Veii

Some progress.
RTT is on auto, but changed SA to 1.23V and TX to 1.33V MC to 1.518V.
I believe I will play with TX and MC further.
1705850374400.png

1705850424694.png
1705850431878.png
1705850439356.png
 
Should I lower the Voltage and increase the RD to 80?
Dont touch voltages :)

If you get RTTs figured out without DFE override, it will be top

RTTs need to be able to run and cover 1.44-1.6v
Actually they should be fine till 1.65 even

Current state on 34-80-60-40-48. It's still running. RX-DFE on auto atm.
I was looking for the detailed comment from you about RTT's, but can't find it. Should be bookmarked somewhere I believe.

Let's see if lowered Dqs fix it:
DQS either one above or one bellow PARK
Dont match it.
Matching will require higher delta on pmic, to get some sort of usable distance between voltages.

NOM shenanigans indeed sounds like some different/odd behaving pcb.
I wonder.

Tuning RTTs on VDD/Q matched (in mem)
May be a good idea
But this will mess too much up.

Ya it looks fine.
You can play around if one NOM needs to be off or on
And you can play around between 40-48 PARK
If that changes how weak one NOM can reach.
Target remains 80-80. Those are not dual sided.
Beitrag automatisch zusammengeführt:

Hi @Veii

Some progress.
RTT is on auto, but changed SA to 1.23V and TX to 1.33V MC to 1.518V.
I believe I will play with TX and MC further.
Sorry, today has been one of those away from home days.

We do want low SA tho.
Can you track how low you can go without y-cruncher loosing and/or rather gaining computational speed ?

Should be like 6* 10 min work at worst.
Move in steps of 5-10mV. Dont move over 10mv Steps
 
Don't touch voltages:)

If you get RTTs figured out without DFE override, it will be top

RTTs need to be able to run and cover 1.44-1.6v
Actually they should be fine till 1.65 even


DQS either one above or one bellow PARK
Don't match it.
Matching will require higher delta on pmic, to get some sort of usable distance between voltages.

NOM shenanigans indeed sounds like some different/odd behaving pcb.
I wonder.

Tuning RTTs on VDD/Q matched (in mem)
May be a good idea
But this will mess up too much.

Ya it looks fine.
You can play around if one NOM needs to be off or on
And you can play around between 40-48 PARK
If that changes how weak one NOM can reach.
Target remains 80-80. Those are not dual sided.
Thank you. I believe the top settings could be the foundation now I will do more with RTT's again.
I start to believe that it could be doable. 🙃
Should I touch the ODT too?
 
@zebra_hun
Ring hits ICCMAX.

No way around undervolt to lower UncoreVID.
Or limiting max ring multiplier.
Hitting PL4 means throttle.

How would it throttle.
At best package throttle, at worst voltage supply throttle

What happens if voltage is throttled too much ~ craaash 🤭
Mostly on cache

EDIT:
Unlimited ICCMAX makes no sense,
As you trade away cpu lifetime for "short time stability"
Fixing curve is the path~
To lower one part of supply, leave more margins till ICCMAX. Leave more margins for IMC.
Beitrag automatisch zusammengeführt:

Thank you. I believe the top settings could be the foundation now I will do more with RTT's again.
I start to believe that it could be doable. 🙃
Should I touch the ODT too?
You can just reuse them.
SA, VDD(2)_CPU, VDDQ_CPU
It either runs or doesnt.

But i would focus to keep SA lower.
VDDQ <-> MEM is an own topic.
Dont mind it for now.

We both know that things are cross connected
Low SA baseline is important. Else other voltages spike up and some lower.
Generally target since recent bioses is strong ODT.

Need to work with it.
Increasing SA is not the path to take.


@mahaudi Karhu 20000%
Es braucht solange es braucht :)
 
Zuletzt bearbeitet:
@zebra_hun
Ring hits ICCMAX.

No way around undervolt to lower UncoreVID.
Or limiting max ring multiplier.
Hitting PL4 means throttle.

How would it throttle.
At best package throttle, at worst voltage supply throttle

What happens if voltage is throttled too much ~ craaash 🤭
Mostly on cache

EDIT:
Unlimited ICCMAX makes no sense,
As you trade away cpu lifetime for "short time stability"
Fixing curve is the path~
To lower one part of supply, leave more margins till ICCMAX. Leave more margins for IMC.
Beitrag automatisch zusammengeführt:


You can just reuse them.
SA, VDD(2)_CPU, VDDQ_CPU
It either runs or doesnt.

But i would focus to keep SA lower.
VDDQ <-> MEM is an own topic.
Dont mind it for now.

We both know that things are cross connected
Low SA baseline is important. Else other voltages spike up and some lower.
Generally target since recent bioses is strong ODT.

Need to work with it.
Increasing SA is not the path to take.


@mahaudi Karhu 20000%
Es braucht solange es braucht :)
I will try to reduce it. Tomorrow will check the current setup with lower SA.
 
Das ist doch ein schönes neues Daily-Setup

8400 CL 34 47 46 32 @ 1.590v

Richtig guter Memory Score @ Geekbench 3......und AIDA64 sieht auch gut aus.

:teufel:
Mit 8400C34 kannst du auch bessere Effi erzeugen :); 8400C36 mit nur P55/R50

IMG_2478.jpeg

Ein bisschen mehr geht auch noch 😇

IMG_4436.jpeg
 
Registered in this forum to try my luck finding tips.
Im currently on Encore with 7800Gskill Adies
I believe in "goal is always to overshoot in defensive side of things instead of always being at 1.0 ratio of cooling/clocks"

Im with barrow aluminium heatsinks with holes on top (great btw, 1.5mm pads) + noctua fan 2000rpm at ram

I did overclock to 8000c36 and 7800c34 but i didnt really gained anything in games and honestly i just want to gain MORE stability
I tried read about skew odts but honestly i am too dumb and dont have an osciloscope to really see it and hate doing guesses.

Im using xmp tweaked with few changes and zero voltage changes, maybe shamino settings is better than guessers in OCN and my guesses.

if anyone has an skew/odt recommendation to make what i have even more stable, thats my goal
any gain that isnt really pushing speeds
even lower vdd/vddq ram i accept any tip

even if no answers , the forum looks great, had alot of good reads.
im from brazil i use translate
edit: this is retrain stable 24h karhu 1h vst/n63/vt3 and performance is good, just wanna see if i can try anything else

i used to read recomendations from pahktunov/vei
but in ddr5 im just lost when it comes to ram powersaving(ppd and etc) and other uncommon settings, and encore have literally all, its depressing not know
 

Anhänge

  • ram.png
    ram.png
    80,7 KB · Aufrufe: 58
Zuletzt bearbeitet:
the forum looks great, had alot of good reads.
success 🤭

If you read all past 10-15 pages
I dont know what to add ~ all is written

Im using xmp tweaked with few changes and zero voltage changes, maybe shamino settings is better than guessers in OCN and my guesses.

if anyone has an skew/odt recommendation to make what i have even more stable, thats my goal
any gain that isnt really pushing speeds
even lower vdd/vddq ram i accept any tip
it just doesnt work that way
Being served on gold spoon - works for one person, doesnt work for other

And you dont learn by reading
But by trying out what you read.

No memory related voltages on auto
no cpu fixed clock, no cpu fixed voltage

I'm sorry if this comes over rude
But question read like , "explain me everything"
Ask question by question, and maybe it can be answered :-)

Every question maybe has long or short explanation.
Its too much work and discouraging when asking "explain me everything".

ODT and RTT depend on mainboard
Depend on dimm
Depend on used voltage ~ which depend on CPU sample
Depend on PLL state.
Depend on hidden Bios settings, changing every 1-2 Bios releases.

I dont know how to explain or answer this.
 
I'm sorry if this comes over rude
But question read like , "explain me everything"

Its very fine.
I dont know what to ask

Thing is i dont know what I am looking for. if there is something to look for

In past i would disable powerdown and selfrefresh, reading a bit aparently disabling it dont even disable it fully (and lowering the timings make pyprime faster so it does something)

and even if i set odts that work, how im supposed to know if my signal is less noisy/more correct

depth of knowledge = money and i dont have 100k$ for osciloscope xd

so maybe hints from ppl that know ppl with it. idk

being able to SEE is the first step to getting results, I honestly dont know how im supposed to know if X or Y is working just by stressing. it seens very guessable

edit: ram/latency is king of optimization so im just trying to make mine better while not pushing any stability boundaries
 
and even if i set odts that work, how im supposed to know if my signal is less noisy/more correct
By pushing boundries and seeing what sticks
One can need lower voltage but still limit in max clock = too noisy
One can need much more voltage, but allow to scale higher than the dimm pcb was designed (example DDR4 , weak A0 PCB)
If you either push much, or start to learn on what many people call "g*rbage tier" or just bad bad cheap gear
You will understand when something "scales well".
reading a bit aparently disabling it dont even disable it fully (and lowering the timings make pyprime faster so it does something)
This is the fun thing i had to learn
Lower is always faster somewhere.
On some application with some dataset size, big or small
You will always find some application where you can say "see it scales"

But thats always the case.
Lower is in specific type of read combinations or read, no-operation, read ~ always faster
some dont even leave one IC.

The question then becomes not about if the timing scales, but what was the tradeoff
What did you influence before set timing an what follows after set timings
Maybe its complicated and sounds like nonsense for somebody who only focuses on the result
But thats how i learned to stop just pushing pushing. And try to learn what happens and what do i pay with my changes

No timing goes alone :)
Absolutely zero.
Minimum has two links , often its 3 (before, target, effect after)
depth of knowledge = money and i dont have 100k$ for osciloscope xd
Money or lifetime :) , or connections
If you dig enough even if nobody wants to help you (take me as the best "worst example")
you still learn. Maybe not in 1month, maybe its difficult for 5 months
But after 6-8 years being silly and fooling around ~ you sloowly start to understand.

I'm still novice, but even if you hold bad of me (example)
I learned a bit. Day by day more. Soo if low-iq me can do it ~ everybody can.
just work harder 🤭 and don't forget your goals
edit: ram/latency is king of optimization so im just trying to make mine better while not pushing any stability boundaries
This might sound naive from me
But it is really not the case :)

Latency matters only to be faster than inter-thread/inner-cpu latency
Soo priority to leave cache and leak to mem, vs other cores ~ is higher.
If memory access is fast enough, decision to pick that path is rather whats going to happen ~ vs make roundtrip and stay on cpu side.

Latency only is a variable vs how the cpu behaves.
Some SKUs scale more, some less, some who use huge L3 cache or fast V-Cache
~ its hard on them to tell to leak to memory, when memory maybe differs between 15-20GB/s, but internally it runs at 500GB/s with same latency . . . for example
^ like why should CPU even bother to leak to mem, instead of wait a bit and process it much faster or split around other free cores~

Many many IF's
Try to get familiar with SiSoftware Sandra.
Don't rely too much on those little memory filling random-type of benchmarks.
They will show one scenario, but memory can be optimized, for all sorts of different workloads.
Its not a simple or straightforward thing.
Beitrag automatisch zusammengeführt:

Latency matters only to be faster than inter-thread/inner-cpu latency
Because how would you define "faster" or "slower"
if you only test in one made up coded scenario.

How can you tell its actually "faster".
How can you tell its not just using one timing more, because the target/workload is just such little one.
Try to look at the bigger picture~

No timing goes alone
Lower Clock = lower visual timings.
but timings are just placeholders. The value is just visual without any meaning
It gets a meaning when you compare it to what type of clock ~ how much, it actually is
And still then - first target is stability high clock.

Timings are just gifts. Extras. They have low priority.
How things interact with each other, is more important.
And of-course absolute stability.
 
Zuletzt bearbeitet:
im really demotivated to deep dive cause it will be guesses..

an pc is faster cause he can do instructions faster
not cause it has been overclocked higher "grain of salt"xd
im more interested in architecture and his awful relationship with scheduler/os

but i came to the right place, the insight fullfilled me on what to do a bit
most overclocks i see put more stress on the arch and is just average fps streetching with barely any gains on lows where it actually improve ur experience

i cannot get access to any good pdf to read (last time i tried some intel pdf i got the text "it's a confidential slidedeck so I'm not too comfortable putting it "out there" without the proper approvals") but still managed to find a bios setting that learn that it was Power Control Unit/throttling that i was able to disable and get 80w more of my cpu at same load, was fun)
and too dumb to get an conclusion from jedec


just one objective question, do you think i am wrong of disabling powerdown and set timings to minimum register
it showed no signs of unstability, improved some benchs(will try what u recommended), and i do that cause of 10y old pdfs that i read about them,
 
just one objective question, do you think i am wrong of disabling powerdown and set timings to minimum register
YES !
Because powerdown "usage" changed

Now powerdown is part of self-refresh
and self refresh is part of FGR. Clock-Halting is also a fun new thing.

FGR especially is mostly used on LPDDR4 (onwards) and you will find information on GDDR5++
Everything LPDDR4 can do, DDR5 can do.
Just DDR5 changed from where many things were under MSR ~ they are now API calls.

DDR5 became a standalone ecosystem similar to your GPU.


Setting timings to lowest point, will harm some timings before and some timings after.
// Because many things - multiple timings happen at the same time !

JEDEC paper i still "don't understand" fully.
But they (memory committee, jedec is there for all sorts behavior) - pads/paste/nand-flash/thermal behavior, pcb layering and targets . . . many things :)
Are the first source you can use as "this is how it will be , if nothing was changed mid or past design" . Especially when DDR5 is still in its young age.
Like "this is the target and this is how it has to be. Change it up a bit , but still this is how it behaves and can behave". The bare minimum required targets for IC design.

On that part you can be sure.
Its like other global comitee's that set the standard and thats how things are and will be.
Its your book to see if you misunderstood something and check if you broke some rule.
Memory will do its best to correct and jedec papers do not define all sort of use-cases

For this you have to explore private researchers
^ couple interesting ones ~ last is complex and becomes interesting on AMDs side.

For example , and similar from https://ieeexplore.ieee.org/
Be it for GDDR or LPDDR or own designed FPGA circuits.
I can guarantee there is always some snippets of information that you can use :)

If you search how to get those papers, it will help.
I can not be the person that helps in going around university restrictions.
Even if as researcher person think knowledge should be free for everyone.

Nvidia has early on employed some tallented Memory Researchers
There might be some information out on this for LPDDR4 and GDDR.
There are parallels. Just the link layout and location layout differs a bit.
Else they are still very similar to consumer ICs.

EDIT:
Another location you can dig is with Baidu search (chinese) *
For example on forums like https://blog.csdn.net/
Or the korean field on forums like https://blog.naver.com/

Its unfortunate that NDA mentality is a thing.
Rivalism and secret-keeping is strong in this industry :( Unfortunately that halts global progress, but money i guess~

* Soo you still need to work a bit for your goal.
Just don't forget your goals ~ and if you learn or be successful, give back and help more people :)
If you don't, you do no change to this "hard to find information" issue. So other people will equally struggle like you did.

EDIT2:
Here are some writeups from one blogger on CSDN [CN]
If you want to figure things out ~ you'll figure things out 🤭
But i've given enough, now its on you to dig and slowly with multiple readings try to understand :giggle:
im really demotivated to deep dive cause it will be guesses..
And not only read and try to understand
But play around.
With imagining to guess, nothing works. With trial and error you will understand
Many people hand out information, but spoonfeeding is not ok.
"gifted" vendor-documents and knowledge are earned, not handed out just like this.
 
Zuletzt bearbeitet:
Yoo,

Moved to Z790 from AM5 and this are bit different here.

VST+VT3, TM5 and karhu stable base. I've made some very basic changes what did work with AM5.

Voltages are not minimized to lowest possible.

Any points for extra tightening?
 

Anhänge

  • base.png
    base.png
    159,3 KB · Aufrufe: 42
There is picture attachment, you cannot see it?
Sorry, I ment the quoted things.
In forums, its not common to work on "trust me" bases.
Looking for TM5 & y-cruncher screenshots.

Are they uploading still ?
 
Hardwareluxx setzt keine externen Werbe- und Tracking-Cookies ein. Auf unserer Webseite finden Sie nur noch Cookies nach berechtigtem Interesse (Art. 6 Abs. 1 Satz 1 lit. f DSGVO) oder eigene funktionelle Cookies. Durch die Nutzung unserer Webseite erklären Sie sich damit einverstanden, dass wir diese Cookies setzen. Mehr Informationen und Möglichkeiten zur Einstellung unserer Cookies finden Sie in unserer Datenschutzerklärung.


Zurück
Oben Unten refresh