[Sammelthread] Samsung 8Gbit DDR4 C-Die K4A8G085WC - OC-Ergebnisse im Startbeitrag! [updated]

[Operation.Void]

I found 1.33V unstable in my case, and had to bring it back to 1.34V and it's been weeks without issue. I also decided I wanted to test CAS vs FCLK. I ran 3733CL18 and 3666CL17 through the Rainbow Six Siege benchmark at 360p on low settings with my 3500X and got the exact same results. I give the win to FCLK because it also helps multithreaded workloads which do not care about RAM.

Update: I have further reduced some timings! It never stops ...

I reduced tFAW from 28 to 27, tRRD_L from 6 to 5 and tRCDWR from 19 to 18.

Update 2: For further proof of stability I have ran HCI Memtest for 100% on each thread, not a long test but eh the real test is proper usage rather than wasting too much time stability testing the OC.

Update 3: I tested overnight with TM5 for 5 hours or 20 cycles with no WHEA errors!

Update 4! Shrunk the previous images as far as they could go to declutter this message. I have further tightened timings, going from tRAS 39 to 36 and getting all my WRWR and RDRD (except SCL) to 1. I suppose having only 2 DIMM slots helps with this. This screenshot here is of 50 minutes of TM5 Anta777 heavy5opt, though I have done more testing on the exact timings. 20 minutes but I was using the computer and as a result TM5 decided to stop due to lack of allocated RAM and never came back, and another hour previously in which I forgot to take a screenshot. So did 18 minutes of HCI. Cumulative 2.5 hours. All with no WHEA errors either. Tested overnight for 6h38m (25 cycles) and got no errors. Raised tRFC to 584, because I don't wan to be on the edge. None of my timings go any lower, none. Technically tRAS does, but lowering it further than 36 yields no gain and trying 32 actually scored worse on AIDA64 latency test.

I also broke the 70 nanosecond barrier with this at 3955MHz on the CPU.

 

[alxns]

Hast du das Bios manuell oder über die MSI APP geflasht ?

M-FLASH mit USB FTW! I wusste nicht dass es über eine APP möglich war. Nicht vergessen zuerst Optimized Defaults laden oder Clear CMOS!

@Operation.Void ...but have you passed the ultimate test for our beloved c-die?
after 8hr of Karhu pass
2hrs of TM5 1 error
I bring you...

AIDA64 30min fail. and that's not even its final form. Usually it would fail in 10min. With 1.33V. 1.34V it will fail before even loading up the memory 😅
OK admittedly running 4.3GHz doesn't help with stability. Funnily enough, best latency yet was with tRP raised to 26. If someone cared to explain.

Anyway, that was still with a beta BIOS, which had some weird USB issues and problems correctly saving settings...
I updated to final AGESA 1006 BIOS (A70), and first try with the same settings apart from VDDG at 1050mV and tRP back to 21, 5hr of Karhu pass


now...

 

[Maurise]

and i can give you settings that aida is fine with but karhu gets at 7%. always funny people think there is a ultimate stress test. wait until you find out about GSAT! karhu and aida64 >1h stable but BSOD at start of GSAT is not that hard. if you want to test voltages try them on prime 95.

if you have better performance with higher timings it can be stability. or there are other timings that get changed with that that you cant even see. i would check with some other benchmark. Aida64 photoworxx is good. gives you 1 number thats very consistent

 

[Operation.Void]

😭 Why must you scare me like this!

I am trying to get my tFAW lower, it passed 40 minutes TM5 at 25 from 27, a good start. The issue is that I need to try AIDA64 and GSAT.

Update: tFAW 20 passed 50 minutes TM5 Anta777 heavy5opt, failed 1 minute of AIDA64! Now it's been 35 minutes at my previous settings, tFAW 27. I will test for an hour.

Update 2: Final changes are tRAS down to 30 and tRCDWR down to 8, not sure if either of those actually did anything.

"Final" timings (these passed the hour of AIDA as above):

Update 3: Of course those were not the "final" timings!
I got tRTP 5 and tWR 10 running for 40 minutes of AIDA64 😬 and it is running TestMem5 1usmus_v3 for 30 minutes so far, and I will then use Anta777 heavy5opt overnight. Many tests!

Update 4: 50 minutes TM5 heavy5opt fail.

New DIMM cooling solution. Having the fan the way it was created a lot of noise, for reasons I do not know so I have used cable ties to hold it over the top of the DIMMs. No zip ties unfortunately but those would be stronger and more elegant. As you may be able to tell I also replaced the rainbow vomit ATX 24 pin power supply

Currently simulating a hot environment by running the A/C at 24 degrees.

 

[alxns]

Hey I like rainbow vomit cables! No confusion when testing PSUs

and i can give you settings that aida is fine with but karhu gets at 7%. always funny people think there is a ultimate stress test. wait until you find out about GSAT! karhu and aida64 >1h stable but BSOD at start of GSAT is not that hard. if you want to test voltages try them on prime 95.

On C-die? Well yeah... e.g. running VDIMM 1.32V, it will make AIDA happier but the other tests will fail fast. I meant AIDA being the ultimate test more as a joke really 😋 But still in my case all of the other tests took much longer, if not without errors (Karhu, GSAT, Prime large, OCCT large, realbench...)

Now, I'm thinking that if AIDA fails so fast, it's because of fabric-related instability, not directly because of the memory. I knew my fabric was on the edge of stability, because of the Bus/Interconnect WHEA warnings. A higher VSoC fixed that. Now, just turning on clock spectrum spreading with FCLK 1900 would make the machine completely unstable, throwing a lot more WHEA Interconnect warnings/fatal errors, or barely even run the BIOS menu. While 1866 is fine with the standard VSoC (just a slight bump in set voltage to get 1.10 at the die)

So I'm now running 1866MHz. Much better, and I also found out that disabling GDM while keeping command rate 1T was fine, though I had to use a ClkDrv strength of 120 Ohm

Tested in addition to TM5: Karhu 20k%, OCCT large AVX2 1hr, AIDA 1hr

Now would it be possible running a stable x19-ratio FCLK? Maybe by adjusting BCLK, if that's the culprit as enabling the spread spectrum has shown. I really have no idea how all the clock dividers/multipliers work.

if you have better performance with higher timings it can be stability. or there are other timings that get changed with that that you cant even see. i would check with some other benchmark. Aida64 photoworxx is good. gives you 1 number thats very consistent

Danke für den Tipp!

 

[Operation.Void]

@alxns

I believe AIDA will pick up any error within the system (probably picks up WHEA errors), even if it is only testing memory whereas other stress testing software will only pick up memory errors.

 

[alxns]

Yep but to first pick up errors, the test has to induce them. A calculation error can happen anywhere in the system. It could be memory, it could be IMC, it could be Data Fabric, it could be cache, it could be cores. But the problem is, a program has no way of telling where it happens.
WHEA warnings are reported by Windows only, since it's the OS and the kernel. You can have AIDA running (or any test for that matter), get a warning in the event viewer, and AIDA will still be running, because the error was "corrected". If it couldn't be corrected (fatal or critical error), it'll shutdown or BSOD before the program has time to complain.

So one program could be better than another in inducing errors for a particular instability in a particular part of the system. And that's why you can't just rely on one test.

 

[Operation.Void]

Aaaah! According to Windows Reliability Monitor it is showing hardware errors with the code 144 (all of them). After some googling it appears to be video card driver, and I am currently running an RX 570. You think it's my memory OC or AMD drivers doing AMD driver things?

 

[kakyoin]

3733 Cl18-20-20-40, 1.28v AIDA64 stable but fail 1usmus tm5,1.30v 1usmus tm5 stable but AIDA64 now fails at 3 minutes. More voltage aida fail, less voltage tm5 fail, Dammit c-die

 

[Operation.Void]

3733 Cl18-20-20-40, 1.28v AIDA64 stable but fail 1usmus tm5,1.30v 1usmus tm5 stable but AIDA64 now fails at 3 minutes. Dammit c-die

Increase tRCDRD and/or voltage, never gotten tRCDRD 20 to work at 3733MT/s. Could also bring tRAS down a fair bit, mine goes down to 30, but theoretically 38 is the minimum before there are no gains.

 

[kakyoin]

Increase tRCDRD and/or voltage, never gotten tRCDRD 20 to work at 3733MT/s. So could bring tRAS down a fair bit, mine goes down to 30, but theoretically 38 is the minimum before there are no gains.

Whoops, it was Cl20-20-20-40 at 3733, not Cl18. At Cl18 3733 1.28v won't even post, while 1.3v fails AIDA as usual... even Cl20-22-22-42 will fail.
I'm running dual rank, 2x16, so it might be the reason why I can't push the clock higher... For now I'm trying 3600 Cl18-21-21-41, tFAW 16 and tRFC 576... tm5 1usmus, aida64, tm5 anta777 stable so far... Let's see how far I can push the secondary timings...

 

[Operation.Void]

Whoops, it was Cl20-20-20-40 at 3733, not Cl18. At Cl18 3733 1.28v won't even post, while 1.3v fails AIDA as usual... even Cl20-22-22-42 will fail.
I'm running dual rank, 2x16, so it might be the reason why I can't push the clock higher... For now I'm trying 3600 Cl18-21-21-41, tFAW 16 and tRFC 576... tm5 1usmus, aida64, tm5 anta777 stable so far...

You're being very conservative with voltage, remember the C-die kits use 1.35V stock so you're safe there. tFAW 16 is amazing, but if I go below 27 it errors after about 30-40 minutes of TM5 1usmus_v3.

 

[kakyoin]

at 1.35v-1.33v 3733 and 3800 the screen will flicker in and out, need less voltage for higher frequencies... but with less voltage will crash in tm5 1usmus...

3600 18-20-20-40 on my kit will crash within 3-5 minutes of tm5 and voltages from 1.3v to 1.35v made zero difference the primary timings are pretty much unresponsive to voltage

 

[Operation.Void]

at 1.35v-1.33v 3733 and 3800 the screen will flicker in and out, need less voltage for higher frequencies... but with less voltage will crash in tm5 1usmus... the primary timings are pretty much unresponsive to voltage

3600 18-20-20-40 on my kit will crash within 3-5 minutes of tm5 and voltages from 1.3v to 1.35v made zero difference

You sure FCLK is stable if on Ryzen? Maybe increase SOC?

 

[kakyoin]

You sure FCLK is stable if on Ryzen? Maybe increase SOC?

flck 1900 is pretty much stable... I tried both stock and 1.15-1.175soc... The flickering only goes away once I drop dram voltage.

 

[Operation.Void]

flck 1900 is pretty much stable... I tried both stock and 1.15-1.175soc... The flickering only goes away once I drop dram voltage.

Lucky! Flickering going away is probably because you have poor cooling and it can't tolerate 1.3-1.35V. Haha 10ns CAS peasant!

 

[kakyoin]

Lucky! Flickering going away is probably because you have poor cooling and it can't tolerate 1.3-1.35V. Haha 10ns CAS peasant!

sad reaccs only... I'm on a very tight mini-ITX case so more cooling for the RAM is out of the question... currently doing tightening timings down, tFAW 16 tm5 1usmus has been stable for 2 hours now, i think i'll move on to the next timings.
I do have another kit of c-dies, 2x16 with the same part number. So I might try OC'ing on that kit to see if there's any difference

 

[Operation.Void]

sad reaccs only... I'm on a very tight mini-ITX case so more cooling for the RAM is out of the question... currently doing tightening timings down, tFAW 16 tm5 1usmus has been stable for 2 hours now, i think i'll move on to the next timings.
I do have another kit of c-dies, 2x16 with the same part number. So I might try OC'ing on that kit to see if there's any difference

Liquid cooling? Lol. 1.28V should do anything at 10ns CAS, C-die doesn't need extra voltage for frequency with same CAS from my testing. Make sure when testing to have a heater on or something (realistic warm room scenario)! At 21c I can pass TM5 for ages but at ambient 24c it gets errors within 2 minutes.

 

[kakyoin]

Liquid cooling? Lol. 1.28V should do anything at 10ns CAS, C-die doesn't need extra voltage for frequency with same CAS from my testing. Make sure when testing to have a heater on or something (realistic warm room scenario)! At 21c I can pass TM5 for ages but at ambient 24c it gets errors within 2 minutes.

Haha, it's very hot right now in my place so if anything, I'm testing at the hottest ambient scenario. From my testing 1.28v is not stable for tm5 at any frequency (3600-3800), I need 1.3v in order to pass tm5. AIDA64 on the other hand needs 1.28v for 3666 or 3733 to pass, higher voltage will fail

 

[kakyoin]

All timings tightened! 3600 CL18-21-21-41 @ 1.3v with very tight secondary timings on my dual rank kit. TM5 1usmus over 1 hour, Aida64 stable, now testing TM5 anta777 overnight.

My conclusions:
- Primary timings are set pretty loose, since my kit negatively scales around 1.28-1.3v ; 3600mhz at 1.3 to 1.35 volts made no difference, while 3733-3800mhz became unstable at 1.35v and needed to be dropped down to 1.28v for AIDA64 stability (tm5 will crash in 3733-3800mhz regardless of voltage). It seems like AIDA64 prefers less voltage while tm5 and other tests prefer more
- Secondary timings are a breeze. Once I've stabilized my primary timings, most if not all secondary timings can be very tightened

Seems like my kit is middle of the road as far as c-die goes.

Hopefully I can test the other c-die kit over this week.

EDIT: got Geardown disabled + Command Rate 1T working, had to use 40-20-20-24 drive strength... currently testing
EDIT2: GD disabled + CR1T with 40-20-20-24 drvstr, 1.31v AIDA64 stable for almost 2 hours, TM5 1usmus overnight stable, tm5 anta777 stable. Pretty decent results.

 

[alxns]

Looks good @kakyoin ! Haven't been testing for a while... I need the computer stable at the moment. 4.3GHz 3733 18-18-21-22-40 1T GDM disabled so far so good, although I wonder how well it would pass tests now

 

[Operation.Void]

Nice @kakyoin. Seems the point of negative scaling up to 1.35V is mostly influenced by temperature, but after 1.35V it's pretty much just C-die doing C-die things. I think you could bring tRAS and tRC down. tRC probably 62 and tRAS 39. Also try bringing ProcODT down, I have heard above 40 is useless and possibly detrimental to Zen 2. 34-38 is the sweet spot, I am running 36.9.

 

[kakyoin]

Nice @kakyoin. Seems the point of negative scaling up to 1.35V is mostly influenced by temperature, but after 1.35V it's pretty much just C-die doing C-die things. I think you could bring tRAS and tRC down. tRC probably 62 and tRAS 39. Also try bringing ProcODT down, I have heard above 40 is useless and possibly detrimental to Zen 2. 34-38 is the sweet spot, I am running 36.9.

tRC is already tight; Setting any tRC lower will make the machine refuse to boot, and tRAS lower than tCL + tRCD + 2 might actually incur a performance penalty. ProcODT I haven't mess with, so I'll probably try that out when i'm not using this machine for work

 

[Operation.Void]

@kakyoin you can always try I got my tRAS as low as 30 without performance penalty.

 

[Erdy]

Hallo zusammen. Bin neu hier beim RAM OC

Bekam am Samstag mein neuen Corsair LPX 16GB 3200 CL16 RAM mit einem Asrock B550M Steel Legened und wollte mich das erste mal am richtigen RAM OC probieren. Hatte mich anfangs wie die meisten mega darüber gefreut, als mir der Thaiphoon Burner B-Dies angezeigt hatte, weil es dazu schon viele Erfahrungswerte gibt. Als ich dann erfahren habe dass es C-Dies sind und der DRAM Calculator for Ryzen keine C-Dies gelistet hat, kam die Ernüchterung alle Timings mehr oder weniger selber herrauszufinden

XMP Profil Settings:

Mit den Standard Settings konnte ich noch nicht mal 3400er Takt stabil halten, selbst mit nicht Haupttimings von 20-22-22-55 oder 1.4V. Nach stundenlangem Testen hab ich herausgefunden dass es am tWRRD 1 lag. Wieso der im Bios so hinterlegt war weiß ich im nachhinein auch nicht. Nachdem tWRRD auf 4 stand lief alles bedeutend besser. Hab erstmal mit den hohen Haupttimings, ein paar veränderten Subtimings und tRFC 750 angefangen die Taktgrenze auszuloten. Kam so bis 3866mhz, mein ziel waren aber 3800mhz. Habe fast alle Änderungen mit einem 3min Aida64 Stresstest kontrolliert. Wie handhabt ihr das kontrollieren der veränderten Werte?
Was mir besoders aufgefallen ist, war einmal halt die Sache mit dem tWRRD Wert und dass mein Kit mit geringerer Spannung besser im Takt skalierte. Zwischen 1.28-1.32V hatte ich meinen Sweetspot. Hatte mit den Subtimings fast gar keine Probleme. Hab die meistens leicht unter den hier durchschnittlich erreichten Werten gesetzt, welche fast immer problemlos durch den Aida64 Stresstest liefen. Also komplett ausgereizt hab ich mein Kit noch nicht komplett, da ich im Grunde mit den Werten sehr zufrieden bin.
Mein jetztiges Problem ist, dass ich nicht mehr den anfangs erreichten FCLK von 1900 schaffe, sondern nur noch 1867mhz. Also nochmal alle Timings optimieren

 

[Kammi]

Auf meinem neuen System ASUS B550 TUF Gaming Plus mit Ryzen 5 3600@4.2
läuft der Billigspeicher Corsair VENGEANCELPX16GB (2x 8GB) DDR4 3600, mit C-dies . Dieser läuft bei XMP-Settings bereits auf 3800/1900 ohne Anpassungen stabil. Laut Aida habe ich dadurch ca. 52.000MB/s Latenz ca. 70ns.
Ich werde die im Thread zusammengetragenen Info's nutzen, meine bislang vergeblichen Bemühungen nochmal zu intensivieren die Timings zu schärfen.
Ich poste bei Gelegenheit mal die Werte.

 

[Erdy]

Wäre sehr interessant für mich mit was für timings die 3600er laufen
Habe es bspw nicht geschafft die tCL auf 18 zu bekommen.

 

[Kammi]

das sind die Profil-Timings

Beitrag automatisch zusammengeführt: 28.07.2020

bei 3766 (rechts) laufen die Riegel mit schärferen Timings auch, bringen aber so keinen Vorteil und verhindern merkwürdiger Weise Warm und Kaltstart?

 

[Erdy]

das sind die Profil-Timings

Beitrag automatisch zusammengeführt: 28.07.2020

Bis auf die tCL liegen fast all meine Timings tiefer, wodurch auch sicher die 18 tCL erreicht wird.
Ich vermute mal dass man den Takt bei dir nicht mehr wirklich Symetrisch steigern kann. Das ist ja schon echt sehr gut so. Aber bei den Timings hast noch echt Potenzial!

[Guide] - RAM Timings und deren Einfluss auf Spiele und Anwendungen (AMD) - Update 23.05.2020

RAM Timings und deren Einfluss auf Spiele und Anwendungen Testsystem | Aida64 Benchmark | Konvertieren und Rendern | Spiele Benchmark | RAM Takt vs Timings | Zusammenfassung Einleitung In diversen Tests wurde bereits festgestellt, dass durch das Optimieren von RAM Timings die Leistung...

www.hardwareluxx.de

In dem Thread sieht man gut welche Timings den meisten Einfluss haben und welche man zuerst ausloten solle

Edit: Habe gerade das Problem dass ich den FCLK nicht zusammen mit dem MCLK auf 3733 bekomme. Erst wo ich die CAD BUS werte wie du auf 24-24-24-24 gestellt habe und GearDown enable auf on gemacht konnte ich beides auf 3733 setzen. Kannst ja mal testen ob dein Setup mit GearDown auf disabled auch stabil weiterhin stabil läuft.

 

[Kammi]

Nein, probiert mit XML-Einstellungen bei 3600