overclocking

[steady lance]

Ohhh, I am blind.

So GDM should be on for 1T?

[dull ginkgo]

GDM is supposed to help make 1t more stable with a bit of performance loss

almost like 1.5t kinda thing

[steady lance]

Small amount I assume?

All the ZenTimings posts I have seen for 6400 2x32 has had GDM on. On AM5.

[steady lance]

Mr1111, does vSOC affect anything for Nitro settings and will it affect anything for the RAM besides frequency?

I dropped it to 1.2v no issues and I think I may be able to do 1.1v. I heard fCLK likes low vSOC Voltages.

6400 1:1 at 1.1v SOC posted.

Not Karhu stable.

Trying 1.15v

1.15v appears to be fine.

Going to test 1.12v soc

[wind willow]

If i tweak my ram. It won’t post

Period. Kotex fits

Other than expo^

[steady lance]

What did you change?

All DDR5 sticks should be able to do 1.4V VDD.

VDDIO the safe limit is 1.4v on 7000 series and safe on 9000 but I heard 1.45v might, emphasis on MIGHT be safe.

1.12v was not Karhu stable.

Attempting 1.13

1.13v errors in Karhu

Trying 1.14v

My plan is to find the stable voltage in increments of 10mv then add a 5mv buffer for reassurance

1.14v errors. Errors go from 4 to 1.

[red terrace]

traditionally GDM is one of those "give a little, get a lot" timings, where letting it stay loosened out (GMD on) will generally lend you a much easier time in tightening out more meaningful timings. i typically just leave it on until i'm done with everything else, then see if i can fiddle with it to make it work with GDM off. most of the time it's not worth the bother of it tbh

i haven't done a ton of AM5 dram OC since i find the boot times to be really frustrating, so i can't speak with a lot of authority here. that's how it was on AM4, and with my somewhat limited AM5 experience it seems somewhat similar here

[steady lance]

Gotcha, still thank you.

That is a much better answer then I have seen across 20 or so reddit threads.

Upping SOC to 1.16v.

1.15 semi-stable.

Worked the first run, failed the second.

[steady lance]

I have learned the boot times are heavily affected by the CPU itself.

Motherboard has some affect but not much.

[dull ginkgo]

Idk anything about nitro, haven't really touched ddr5 and never went too deep on ddr4

[steady lance]

This is quite strange.

Same speed/timings at

• 1.15v (SOC), the Karhu speeds are around 267k/269k (AVG/PEAK).

• 1.16v (SOC), the Karhu speeds are around 264K/265K (AVG/PEAK)

Any thoughts on why?

@dull ginkgo @short blade @red terrace

Note that 1.15v was not fully stable, it ran the test fine the first round but failed the second round.

[dull ginkgo]

Run to run variation seems most likely for something that small lol

That's 1% difference

[steady lance]

Posting to keep track.

[steady lance]

note that, with 1.16 the peak and avg climbed very closely together whereas 1.15v it hit peaks but the avg speed was consistently off by 2k.

could definitely still be run to run

but it was something I consistently oberserved over two 10m runs for each.

It feels like 1.16v is the required voltage for 6400 to run smoothly.

Going to try fclk 2133

[dull ginkgo]

I mean if it's not fully stable it could be some random auto timing or voltage running lower/higher than it should be, causing the peaks to be higher, but since it's unstable, whenever there's an error, it brings down the average performance

[steady lance]

1.16v is probably the stable point then.

The avg and peak speed are rising almost perfectly in sync.

1k variance without sudden spikes.

Interestingly Karhu speed does increase with FCLK.

So that makes this a lot quicker to test.

Going to try FCLK 2000.

[wind willow]

Fclk 2000 is standard

[steady lance]

Typo

Meant fCLK 2200

Failed after 50m in Karhu. Going to raise SOC to 1.165 because I think it was the frequency that failed.

[wind willow]

Try 1.2soc

[steady lance]

Doesn't higher SOC make it harder to run high fCLK?

Currently running 1.165v SOC which is showing a 2k bump in Karhu performance with 2200 fCLK vs 1.16v

266k vs 268k avg run speed.

268 bordering 269.

@calm grail haven't seen you here in a while. How are you?

[calm grail]

Good lol

I almost accidentally pinged someone

[steady lance]

I see.

I am like 0.1k away from hitting 269k avg run speed in Karhu.

@wind willow @red terrace thoughts on if I can get above 2200 fCLK stable?

I just hit 269k avg run speed!

[red terrace]

would be pretty cool

[wind willow]

Dont push your luck. There’s hardly any benefit anyways

[steady lance]

269.1 to be specific

One thing to know about me.

[wind willow]

2200fclk is easy with 9950x3d

[steady lance]

If there's any gain to be had, I am going to try

[wind willow]

Upgrade your 3070 and get an asus apex mobo and problem solved

[steady lance]

Mushroom challenged me to hit 300K peak speed in Karhu

So I am almost there. Started at 248K avg.

[wind willow]

Send me a link on something actually readable regarding core cycler for maxing CO

[steady lance]

I can just send you my CC config and tell you where to change the allocated cores + recommend pattern. I learned from Fal'cie, FlexibleFelix, and TrueBacca for my CC knowledge vs reddit.

[wind willow]

And what program can I use to see what my die is for my ram? Just cpu-z?

[red terrace]

nothing reliable. the closest option is thaiphoon burner, but it just reads the SPD, which may contain inaccurate information

[steady lance]

Software is unreliable for detecting dies other than the brand.

[red terrace]

the only truly reliable method is to read the labeling on your ICs

[steady lance]

Wait 10m for me to do a quick VST3 run to check performance on fCLK then I will send the file with some screenshots over for CoreCycler

[steady lance]

I am debating taking off my ram's heatspreaders to check the exact die and upgrading the thermal solution to PTM7950. Thoughts? It's a Kingston Fury kit so it uses screws.

[wind willow]

Use kryosheets

[steady lance]

I know they are Hynix but that's all I truly know based on the label on the RAM kit.

[wind willow]

They’re god tier

[red terrace]

ptm is a bit extreme but you would certainly be able to read the labeling at leaat

[steady lance]

How reliable are those Krypsheets? I am pumping a lot of voltage through my RAM so the thought of it being conductive scares the life out of me.

[red terrace]

the semiconductor part of dram is very small and embedded fairly deep in the package. it's a nonissue

[steady lance]

Gotcha, I have never removed the heatspreaders before

[red terrace]

there is a first time for everything. it shouldn't be too difficult

but best of luck nevertheless i suppose

i have to get to sleep now but i wish you well

[steady lance]

My temps currently peak at 52c with a sustained of 50.3c according to HWinfo after overnight runs of Karhu. Note that I have a 180mm fan blowing onto the dimms.

Good night.

2m left till test is done.

No regression at 2200 fCLK.

@wind willow, see the highlighted blue? That is your core test order, it follows the same numbering system as the bios does.

CCD0 = cores 0-7
CCD1 = cores 8-15

To get the highest peak frequency it is best to set it up to alternate between the CCDs to allow the other one to cool off a little between core runs.

I already have it configured for a 9950x3D to test all 16 cores and here is the zip file for my CoreCycler which has the optimal settings in place already.

If you already have CoreCycler, then just use this config.

@wind willow

Sorry, LordFrog, I accidentally pinged you when I was trying to ping, iLuv.

Okay, for the fun of it I am going to try fCLK 2300

[wind willow]

Ty

[steady lance]

Well it doesn't seem 2300 fCLK is gonna want to boot.

Trying 2267 fCLK next

@short blade I have been having a hard time finding info regarding Nitro settings. I found the Nitro RX and TX but have no idea what to set them to for optimal performance or the risks. Mind giving me some advice?

2267 fCLK is no post. Trying 2233 fCLK next.

[wind willow]

@steady lance Isn't working.

Errors out

I moved the config ini to my folder

[steady lance]

Let me test it on my CC folder give me a sec.

Need to reboot the PC.

2233 fCLK did not post.

[steady lance]

Are you running it with Administrator privileges?

[wind willow]

Ill recheck

[steady lance]

If not it will errors out.

[wind willow]

also how long is this supposed to take

reading this it seems like it takes hours per core

that's not okay.

[steady lance]

CC only uses one core at a time so you can go about your day with it in the background.

[wind willow]

That config file isn't working

[steady lance]

Each cycle per core is about 10m. I usually let it go at least 10 cycles per core to be a extra safe.

[wind willow]

Even with admin mode.

[steady lance]

Try installing my CC zip then. Guessing your version is missing some of the profiles my mentors gave me.

It has every profile you'll ever need for CC and tests the cores with the full suite.

@wind willow

[steady lance]

.zip

[wind willow]

I got it working now so far.

I just copied and pasted the contents inside.

nvm. now it's not following even the testing order.

[steady lance]

@wind willow so this is using my CC folder with my config.

That's what the result should look like.

[wind willow]

lol mine looks nothing like that

[steady lance]

Just download my zip then.

[wind willow]

How old is your file?

[steady lance]

It's a little older but has a lot more testing profiles.

Yours is v11 mind is v9 with modifications.

[wind willow]

kk

[steady lance]

I will download the newer version and test it myself to see if its better when I get around to CPU oc. till then my recommendation is use mine since I know it works.

If your insistent on using the newer version, look for the highlighted blue that I mentioned for setting the core test order.

[wind willow]

I did

Wasn't following the logic.

The program that is

Yours is working in admin mode.

I guess now I need to setup just a CO and see what happens.

instead of all of my CS shenanigans

[steady lance]

To make sure I am following. You downloaded my CC or your CC?

[wind willow]

Both.

[steady lance]

Gotcha

[wind willow]

Your config doesn't work with the newer one

and it does work with the old one in admin mode.

[steady lance]

Mine needs to be run in Admin to make sure it doesn't get finicky with some of the profiles.

[wind willow]

Will it stop when it's done?

[steady lance]

You can set a iteration limit though I personally just close the program or do Ctrl + c when I am done with it.

[wind willow]

If I set to 1 iteration, will it run through all cores and all programs once and then stop?

[steady lance]

Also as said since only one core is being used it will barely impact day to day stuff except for occasional very quick stutters in games.

[steady lance]

If I recall correctly yes. I recommend a minimum of 10 iterations.

[wind willow]

kk

[steady lance]

It will tell you which core errors in bright colors.

[wind willow]

Got you

Will it somehow adjust co for you on the fly?

I doubt it, but you never know.

[steady lance]

It should look like this if no errors come up

[steady lance]

It does not.

[wind willow]

So in that photo core 7 had an error

[steady lance]

No errors, the person did Ctrl + c to Terminate (stop the program)

[wind willow]

oh okay

it just terminated

[steady lance]

Error would look like this.

[wind willow]

got it

ty

[steady lance]

ran another 10m round of Karhu to check my sanity regarding fCLK.

No errors, speed difference was less then 0.01% so well within margin of error.

[wind willow]

now im getting pissed

no matter what I do, this program will do the first core in the setup for a few seconds, then switch to core 11.

And it's when it opens y-cruncher

[steady lance]

It should be doing core 0, then 8, 1, 9.

[wind willow]

nop

[steady lance]

Can you send a screenshot?

[wind willow]

now its randomly testing core 13

That's a fresh start btw

now its back on core 11

[steady lance]

Read the CC info not the HWinfo.

[wind willow]

even task manager is showing that core 11 is being tested

not 0

[steady lance]

CC is set to suspend cores randomly to test for idle/boost stability.

Give it the full 10m without messing with anything.

[wind willow]

k

[steady lance]

If you really want to test, set core 0 to a negative co that you know posts but is not stable.

[wind willow]

I mean, ideally I set all cores to -30 and go from there

[steady lance]

Try it.

[wind willow]

I just want to know who its not following the core order as given in the instructions.

And why its hopping back and forth between core 13 and 11 when doing these tests.

I wonder if its because my cppc is set to frequency

[steady lance]

Likely windows and other apps causing the other cores to spike

I luv any chance you know what are safe VDDG IOD and CCD Voltages and how the correlate to fCLK?

[wind willow]

It didn't even test core 0

lmao

40mV below vsoc

don't put vsoc over 1.3

[steady lance]

I was wondering about VDDG

[wind willow]

Yeah.. This has to be because my cppc is set to frequency

Even OCCT is doing the same thing.

[steady lance]

Try disabling it.

I mean setting it to auto.

[wind willow]

Just did

Now running the tests

nope

that didn't fix it

oh

lol

i know what it is

process lasso

lmfaol

[steady lance]

....

[wind willow]

let me just turn that off

hey look at that

its working now

xD

even with cppc set as frequency

[steady lance]

I am debating if freaking 2233 fCLK is possible without bumping VDDG to the moon.

I am considering moving onto Nitro settings.

[wind willow]

i can't find info on safe vddg

[steady lance]

My board says 0.97v is low and 1.049v is high

I tried 1v debating 1.049

Going to keep fCLK at 2200 till I have better information

I am not planning on frying an $800 chip

[short blade]

1/2/0, burst length 8x/8x

[steady lance]

Just tried 1/2/1 from a reddit post. I am in windows and running Karhu rn. Should I reboot with 1/2/0?

[short blade]

if 1/2/0 is stable then run 1/2/0

if not then run 1/2/1

[steady lance]

K let me try that.

[short blade]

i'm on 1/2/1 because 1/2/0 wasn't stable

[steady lance]

Btw what do you think? 6400 1:1 with fCLK 2200 appears to be stable so far at 1.165v SOC.

@short blade

[short blade]

much better sample than mine

[steady lance]

I checked for performance regression using VST3 and Karhu

[short blade]

i need 1.295 vsoc for 6400 1:1 with nitro 1/2/1

i have 2200 fclk stable at 0.96 vddg

so definitely shouldn't need to go as far as 1.05

[steady lance]

I am not sure what my VDDG is at.

I want to try getting 2233 or higher fCLK to work.

@short blade

Set

• Nitro RX Burst Length 8x
• Nitro TX Burst Length 8x

?

[short blade]

yeah or you can just leave them auto

they auto to 8x

[steady lance]

Will set 8x just to be safe

[short blade]

going lower doesn't seem to impact performance meaningfully but does impact stability

[steady lance]

For VDDG there is an option for IOD and CCD. Which one helps stabilize higher fCLK?

Running Karhu now with Nitros at 1/2/0 6400 1:1 with 2200 fCLK
@short blade

[short blade]

idk ive always set them to the same

good luck

[steady lance]

Thank you and I have another question. I keep seeing reddit posts talking about matching PHY something. What is that about?

@short blade error in Karhu. What stabilizes nitro?

Going to try bumping up SOC to 1.175

I still have my VDDP at 1.15v. Besides frequency what else does it stabilize / destabilize?

@short blade if I had to choose between higher fCLK or tighter Nitro which should I do?

Trying with fCLK at 2000 and SOC at 1.185 to figure out at which voltage nitro 1/2/0 is stable and to measure the performance gain

[short blade]

higher vsoc

[steady lance]

Sustained and peak speed with Nitro 1/2/0 is about 1.5-2 at fCLK 2000.

What does Nitro do exactly and is Karhu Speed a reliable way to check the performance gain?

Karhu Speed 268

• 6400 1:1
• fCLK 2000
• Nitro 1/2/0
• SOC 1.185v

Karhu Speed 270

• 6400 1:1
• fCLK 2200
• Nitro 2/3/1
• SOC 1.165

I am going to try Nitro 1/2/0 with fCLK 2200 at 1.185v SOC to see if that will work.

Errored. I am going to try 1.2v SOC just for the sake of it.

Just hit 271.6 in Karhu. Previous highest was 270.2.

@wind willow when you get a chance, send me a picture of the number directly above the bar code located on the label to the side of your RAM.

Like in the picture above so I or someone else can help verify what die you have.

@short blade 1/2/0 2200fCLK 6400 1:1 not erroring in karhu.

Previous attempts errored out within 60s.

[short blade]

way too early to call it stable but good luck

[steady lance]

fair, but progress.

I plan to do a overnight run of Karhu tonight.

[short blade]

I'm jealous of your IMC

[steady lance]

going to do a few VST3 runs then test fclk 2133 to double check for performance regression.

[steady lance]

I appreciate the compliment. only took two years and two faulty cpus to get AMD to send me something good lmao.

[steady lance]

What would you say is a good Karhu % to call it stable when I do the overnight run?

[short blade]

well the latest I've ever seen an error at was ~25600%

so probably like 30k for me personally

[steady lance]

damn

doesn't that take like 30 hours?

[short blade]

no

that was like 9 hours or so

well yours might take longer cause 64gb

[steady lance]

OHH nevermind... I found my overnight test from last night.

[steady lance]

seems it will take me about 15 hours give or take.

[short blade]

sounds about right

[steady lance]

Dual rank is such a blessing and a curse.

[short blade]

you have significantly higher test speed but double the capacity to cover

[steady lance]

test speed does it scale with capacity in Karhu?

Something I was curious about.

Trying

• 6400 1:1
• Nitro 1/2/0
• fCLK 2133
• SOC 1.9v

Going to run 10m in Karhu then 5 iterations in VST3 to check for fCLK performance whilst increasing it till I am back to 2200.

[steady lance]

I didn't realize VT3 would close if I stopped it. VT3 did 1.37, 1.37, 1.37, 1.36, 1.38 all 10*10.

[steady lance]

I am going to lower soc to 1.195v as test speed did not increase.

[steady lance]

1.195v SOC for Nitro 1/2/0 with 2200fCLK has yielded the best results so far.

I am going to try raising VDDG voltages to 1.05v to see if performance increases or decreases.

Applied 1v VDDG-CCD and 1.05v VDDG-IOD now running Karhu and for a short burst I saw 287K test speed.

295k is the new Max I witnessed during Karhu startup but not the max for sustained during the run.

Max sustained is 273.9K currently.

[steady lance]

@short blade Most stable and best run yet.

VDDG-CCD = 1v
VDDG-IOD = 1.05v

I am debating trying a lower vSOC since increasing VDDG appears to have helped reduce the necessary vSOC needed.

VST3 variance dropped from 2 to 1 and increased sustained + much increased max Karhu speed.

[short blade]

i don't think there's such a thing as vddg reducing required variance

i think you are just drawing conclusions too quickly

i test every change for a few hours minimum

[steady lance]

Some reddit threads suggested that raising VDDG could help stabilize fCLK so that's why I tried it.

Definitely will require a overnight test to be certain I agree.

[steady lance]

Has Hynix released their 3rd generation (B) die yet?

I downloaded Hynix's decoder PDF for DDR5 and saw a list of RAM generations (for example A and M).

[steady lance]

I have work I need to do so I am going to call it there tonight for ram tuning, will run a overnight test before I sleep.

Best and most stable run today.

VDDG-CCD 1v
VDDG-IOD 1.05v

[steady lance]

Welp unfortunately it was not 100% stable.

[steady lance]

Raised VDDG by 0.05v

[steady lance]

What is the fastest way to test fCLK?

Also is there any benefit to adjusting ProcODT PullDown? I already have ProcODT PullUP fairly dialed in at 30ohms.

@short blade what is the difference between Nitro 1/2/0 vs 1/2/1? I recognize that 1/2/0 should be better but I have no idea what it does other than it having something to do with the CPU.

retesting fCLK 2133, 2167, 2200 with the raised VDDG to check for any changes.

fCLK 2167

[steady lance]

fCLK 2200 so far is consistently showing higher peak and sustained speeds versus 2133 or 2167.

[steady lance]

Since Karhu let's me set the Auto Test Size, can I reduce it to say 90% leaving me with enough memory to be able to browse and go about my work day with it running in the background?

I am wondering if this idea would negatively impact the test's ability to check for errors.

Because if I can do that, then I can have a stress test running for about 14-16 hours vs 1-2 at 95%.

@short blade I learned something interesting about Karhu. Karhu Coverage appears to scale exponentially faster with lower tCL.

Every thing the same except tCL and VDD for a 10m run.

Approximate coverages
tCL 30/1.6VDD = 315%
tCL 26/1.65VDD = 17600%

Test Speed (AVG/PEAK)
tCL 30/1.6VDD = 275/280
tCL 26/1.65VDD = 294/294

[short blade]

nope

once again drawing conclusions too quickly

[steady lance]

Is something else affecting it?

[short blade]

test speed takes over an hour to level off

even then it doesn't fully level off

[steady lance]

Why does my Karhu overnight runs vary by 1k at most? I have done a 5 7 and 13hr run

[short blade]

once you're past like 2-3 hour mark it's barely going up anymore

[steady lance]

Still why does dropping tCL to 26 vs 30 cause such a massive increase in coverage and over 10k in test speed?

20m in at tCL 30 the most I can cover is maybe about 700%.

14m in at tCL 26 I am at 25,000% coverage but only a 20k increase to test speed.

Something is scaling here but I do not know what.

[short blade]

um

yeah something is very bugged if you're hitting 25k coverage in 14 min

dunno if it's something to do with your system timer getting messed up or what

but I would not consider that a valid result

[steady lance]

Try restarting and see if it happens again?

No corrupt files in windows and no visual artifacting detected via the usual means of moving apps across each other. Or background artifacts.

Stopped the test, here is a much clearer image of what I was pointing out.

going to try restarting to see if it happens again.

[steady lance]

What exactly is Coverage in Karhu?

[short blade]

how much of the ram you've actually tested

so basically it's saying you've tested 19TB of RAM in 16 minutes

so your test speed would be around 18847 MB/s if that coverage was correct

[steady lance]

Gotcha.

So every 100% is equal to my total RAM capacity tested?

[short blade]

oh wait

why is your test size set to 1GB?

that's probably why the coverage is so messed up

[short blade]

well it's 100% of your test size

but yeah

[steady lance]

It was???

[short blade]

leave test size on auto

[steady lance]

Oh lmfao

I didn't even notice.

Thank you.

Now it's scaling properly.

I wanted to ask, I noticed you said that test speed takes an hour or so to essentially stop going up.

Why is it when I run multi hour long tests in Karhu that my test speed (the one you can see moving up/down constantly) only varies from the short 10m tests by maybe 1.5k tops?

[fervent talon]

(the name change is to allow software tweaking as well)

[steady lance]

Oh gotcha.

I was confused for a moment.

[short blade]

dunno mine grows by much more than that

maybe a chip difference

[steady lance]

gotcha, wondering because I have done 4 overnight runs total thus far, shortest (3hr), longest (13hr) and the variance is so freakishly small. Maybe it's being affected by temps?

[short blade]

dunno, I'm in an MFF case with no ram fan

temps steady at ~50c

[steady lance]

My temps never exceed 51c, and is sustained at 49 for long periods at 1.65v VDD. Granted I have a 180mm fan at full speed blasting at it and my ambient is sub 20c.

[short blade]

I have nothing but indirect case airflow

ambient is probably ~30c

[steady lance]

Due to my variance being so low in long runs, I have found over 10+ tests at each speed in 10m runs that rebooting into windows causes more variance than the duration lmao.

By the way, I noticed when I raised VDDG by 0.5 each, my Karhu Test speed went up by about 4k on average with same settings.

VDDG-IOD = 1.1v
VDDG-CCD = 1.05v

where I am currently at for test speed.

boot to boot variance is about 0.5k

[short blade]

probably means your fclk was retransmitting more on the lower voltage

[steady lance]

What does that mean?

Also I just applied some of Buildzoid's 2x32GB Hynix timings, time to see if they work.

[steady lance]

Gained about 22k from applying BZ's general Hynix timings.

errored. going to lower fCLK to 2133 then let the test run for a couple hours while I go shopping.

[steady lance]

I have come to understand what BZ was saying when he mentioned that fCLK does not yield a large performance gain. fCLK 2133 vs 2200 is only about 1.5k difference consistently.

[short blade]

unstable fclk can retransmit to avoid erroring, but with performance regression

[steady lance]

Gotcha.

By the way, 3 hours in Karhu no errors so far. fCLK 2133 speed is 298k.

Is tRFC2 or tRFCsb used on Ryzen 9000?

[short blade]

nope

I set them to 420/69 for fun

[steady lance]

I like it.

I am probably going to do 420 888

@short blade do you have any knowledge regarding which 7900xtx models are better for overclocking?

I am saving up $1,000 to grab a GPU ideally by end of next month.

[short blade]

no clue, sorry

[steady lance]

@short blade what do you think of my progress so far? 298k in Karhu and 1.54^10 in VST3.

[steady lance]

@red terrace ?

Going to try a quick run with GDM 1/2/1 and fCLK 2200 at 1.165v SOC as I recall fCLK 2200 being fairly stable at that voltage.

Not worth it. Speed is lower.

Going to stick with fCLK at 2133 for the remainder of the tests until all timings at fine tuned.

Going to run 1/2/0 with fCLK 2133 at 1.165v while I go eat.

[steady lance]

Nitro 1/2/0 with fCLK 2133 @ 1.165v

30mv less but test speed dropped to 1.54^10 in VST3 compared to 1.55^10 and Karhu speed decreased by 0.5k in peak/avg speed. I feel as though the difference is minor enough to be up to boot to boot variance and that I have been running tests back to back the entire day.

[steady lance]

1.165 was not enough

[steady lance]

bumped it back to 1.195v and it was game stable

[steady lance]

Average fps went from around low 80fps to a consistent capped 120fps with very occasional drops.

[wind willow]

@steady lance any idea what this is

[steady lance]

Not certain but it looks like it's the result of failure during the initialization of one of the profiles with that Core, in this case being SFT.

[steady lance]

In any case it would be due to a unstable oc.

[steady lance]

A bit of advice I have for tuning CO is do it per core so set core 0 to negative so and so then run CC on just that Core for 12hrs then run a All Core Stress Test for a bit at least an hour personally I do 2-4.

For the CO start with values that fail then increment towards stabilize to save time.

@wind willow

Do that and you should have all 16 cores done in about 10 days if you keep at it.

[wind willow]

I am just gonna do -20/-15 and call it a day.

Also considering disabling pbo altogether and just setting manual limits in amd overclocking.

[steady lance]

@short blade so would you consider this stable? Note, I have some tabs I needed to keep open for work in chrome so the speed ended up lower by about 4k during the test but still it's a 13 almost 14 hour long test.\

[short blade]

I would say that's enough karhu, prob run some other tests too though

[steady lance]

Ran VST for about 3 hours with no issues.

[short blade]

aida64 stress test is a good one too

[steady lance]

No license.

[steady lance]

Here's a picture of what I have.

Minus Aida64 because that is long expired.

[short blade]

i feel like you can use the stress test with the trial version

[steady lance]

I activated the trial a while back, tried opening it last night and it said my trial ran out. So do I have to pay for it or can I just reinstall?

[steady lance]

@short blade uhhm, what's going on here???

I never applied a clock override.

[short blade]

just a misread

[steady lance]

That's a lot of misreads lmao.

Also I am surprised, misreads happen with Effective Clock? I thought Effective Clock was the speed measured over a larger period of time.

[steady lance]

going to run TM5, still can't close my work stuff so it will have to be with chrome open.

good night and be back in a few.

[red terrace]

iirc it's clock speed x CPU usage or something similar

it's supposed to measure the 'effective' speed of the CPU based on how much work it's doing vs just what clock it's at

[short blade]

dunno what the cause is but I can definitely tell you that your cpu is not actually boosting to 8.2ghz at stock

[steady lance]

Passed TM5 (Anta777 Extreme). RAM temps peaked at 53.5 on both DIMMs.

Running VST3 for an hour or two.

[steady lance]

@short blade I think it's safe to call my current tune stable.

Passed:

• Karhu (13h30m)
• TestMem5 (2hr at 1.6x32)
• VST3 (1h50m and a separate 3hr run)

VST3 speed is also extremely stable

I will further tighten the timings now that I have a good baseline.

[short blade]

nice, looks good

[steady lance]

@short blade while I was away, I ran OCCT as well.

No errors.

[steady lance]

Going to save the profile and start tightening some timings.

[red terrace]

i get this too but usually with sys info on 3Dmark. bc yeah my 5800x is not boosting to 12.6ghz actually

[steady lance]

It definitely is! Jk

@red terrace, @short blade, is it possible to drop tRDRDscl below 4 on 7000/9000 series Ryzen? Also does tWRWRscl need to match tRDRDscl or can it go lower?

So far tRDRDscl 4 posts and appears to be stable so far but SCL 3 and below are no post.

I am trying to get my Tertiaries timings tuned in tonight then I will do Secondaries tomorrow.

Damn, seems tRDRDSD 6 and tRDRDDD 6 is the lowest I can go.

tRDRDsd and tRDRDdd at 6 I know is stable as it is part of the baseline I tested for the past week. I am surprised I found the lowest stable first try.

[short blade]

unlikely

from what i see, typically 4/4 works with GDM enabled and 5/5 with GDM disabled

some people are able to do 4/4 with GDM disabled but it's silicon lottery

i would suggest your next step is to try to disable GDM

[short blade]

honestly my experience on AM5 has been that many of the timings are stable at the lowest value that will post

or if not stable will error very quickly

not common to see super intermittent errors from timings

intermittent errors much more likely to be caused by voltages

i don't know if that's generally the case, just anecdotal from me

[short blade]

since you have SCLs at 4/4, don't be surprised if GDM off doesn't post - if it doesn't post try SCLs 5/5 first

this is my current daily setup

[steady lance]

If I have any trouble with getting GDM to work disabled could you give me some tips on how to stabilize it?

[red terrace]

on DDR4 i found adjusting termination resistance helped a lot, idk if that carries over on DDR5. never got to into DDR5 OC as i said before

[steady lance]

I forgot about that.

As for termination resistance I have done ProcODT Pull Up but not Pull Down.

Tertiaries from last night appear to be stable

[steady lance]

I feel it is safe to call my current tertiaries pretty darn stable.

I will try disabling GDM.

[steady lance]

Oh awesome, you managed to get 6400 to work stably?

Gosh it's struggling to train with GDM off.

On it's 3rd cycle now

[steady lance]

This is really starting to annoy me.

I loosened scls as high as 8 and upped VDD/VDDQ/VDDIO but GDM off still refuses to post.

[steady lance]

Well I got GDM to finally post to windows but immediately crashes when running Karhu.

Loosing tCL, tRDWR and tWRRD to see if that makes a difference to stability.

.... GDM off is giving me more headaches then trying to stabilize 6600 1:1.

I could really use some help. Reddit is being completely useless as far as information goes and what little I find and try appear to be sketchy.

Tried overclockers site as well.

Will raising my SOC voltage help stabilize GDM off and or tighter timings?

[steady lance]

ran VST3 for a while before I needed to pack my pc. moving homes.

[short blade]

yeah been daily driving it for a while now

[short blade]

uhh as far as I can remember it was just SCLs, maybe vsoc? I tried gdm off pretty much as soon as I determined freq so very early in the process

don't stress too much if you can't run gdm off, it's probably quite a bit harder on dual rank

it's not the most important thing either

6400 1:1 GDM on with nitro 1/2/0 at vsoc that low is very good especially for dual rank

[short blade]

hmmm... I would try loosening tRC

I don't know how accurate this is but I've heard that GDM on helps "mask" some instability

my tRC isn't stable below 60, 48 seems way too low

I did see other people running 48 stable but when I tried it it would error pretty much instantly

I didn't pay attention to their GDM but maybe it was on

[steady lance]

Gotcha, I will try loosening it when my PC is back on.

[steady lance]

I wish I didn't need more than 32GB sometimes lmao. I am like 99.99% confident that this chip can do 6600 1:1 with single rank.

[short blade]

if it makes you feel any better, dual rank does outperform single rank at same settings

and is harder to push to same settings

so you're not "wasting" the good IMC just by going dual rank

[steady lance]

It does. Thank you. I genuinely considered buying a 2x24GB kit lmao

[steady lance]

@short blade any idea why my CPU appears to be power limiting itself?

[steady lance]

Prior to moving, my PC consistently scored +300K in Karhu and draws 190W.

After moving, scores are as low as 213K and power draw is only about 140W.

PBO was set to have a PPT limit of 300W, and I tried changing it to motherboard limits. Power draw remained the same.

Cleared the bios and reapplied the settings.

Hoping that fixes it

[short blade]

check if it reset karhu settings

sometimes happens randomly

sounds like fpu stress got disabled

[steady lance]

What were they again?

FPU Enabled and Cache Enabled?

Solved!

Yeah the settings reset.

Thank you very much.

[steady lance]

Loosened scls and tRC, no luck getting GDM off to post.

[steady lance]

Hmm I found out that my RAM becomes unstable around 55c.

Transitioning to a new home and as a result I am staying in a airstream currently which gets really hot unless I turn on the AC.

No AC, RAM hit 55+ Celsius and threw errors. AC, RAM is now back in the high 40s around 48-51c.

I also decided to keep GDM on, I did some poking about online and apparently GDM off requires a lot more voltage so I tried 1.75v the other day which did work in that I got into windows but it was not stable and my RAM temps were way to freakishly high.

I'd rather be able to have very tight timings and good temps with GDM on the go for GDM off with looser timings and requiring a Delta fan lmao.

.

So I am going to start working on my Secondaries.

Is there a minimum for tFAW?

[steady lance]

tWTRs = 2 from 4
tWTRl = 12 from 16 (8 did not post)

That's my current progress on Secondaries.

Going to try tWTRl = 10.

10 is no luck

Moving onto tFAW

[steady lance]

Trying tRRDs = 4 from 8.

[steady lance]

tFAW is currently at 16. Correct me if I am wrong but tFAW below 16 does nothing.

[steady lance]

played 2 hours on rivals without problems.

[steady lance]

My current secondary timings:

• tRP = 12
• tRRDS = 4
• tRRDL = 6
• tWTRS = 2
• TWTRL = 12

@short blade what's the best way to test tRFC?

I am going to try tightening that next

[short blade]

same way as everything else

[steady lance]

Are these VDDG and VDDP Voltages safe to daily?

I am currently running them on a 9950x3D and X670E Gaming WiFi Plus motherboard. I don't want to keep said Voltages if they aren't safe for daily.

@bench.marc
@short blade
@dull ginkgo
@fervent talon

• VDDG-CCD = 1.05v
• VDDG-IOD = 1.1v
• VDDP = 1.15v

[short blade]

do the VDDGs need to be that high?

i'm running 2200 FCLK at 0.96

[steady lance]

I am not sure, I had raised them when I was trying to get 2233 fCLK stable but haven't tried to lower them yet.

Is it bad to run it that high? If so I will start trying to reduce it immediately.

[steady lance]

.
So I have been mostly following this guide for RAM, based on it is there anyway I can tighten my Secondaries further?

I am not sure if some of the formulas they are suggesting holds water so some thoughts from you guys would be much appreciated

https://www.reddit.com/r/overclocking/comments/1k3o7qe/am5_ddr5_tuning_cheat_sheet_observations_and_notes/

According to the guide:
tFAW = tRRDs x 4
tWTRS = tRRDS / 2
tWTRL = tRRDL x 2

My current timings

If what the guide says is true, than my secondaries are as low as they will get, but I know that all my secondary timings except tRRDL can post at much lower settings.

.

I have set VDDG to auto and fCLK to 2000. I am going to finish tuning my timings then revisit fCLK and VDDG Voltages.

[steady lance]

Making one more attempt at GDM off.

Using looser timings, higher vSOC and higher VDD.

No post.

Going to try GDM off with XMP timings

I will be damned if this doesn't post

Good it posted.

Screen froze when staring Karhu.

[steady lance]

@short blade thoughts? I feel like something is keeping me from running GDM off even with very VERY loose timings.

[short blade]

what's the zentimings

[steady lance]

Every thing set to auto, same Voltages. GDM off will post.

[steady lance]

32-38-38-38-80-112 for primaries.

Rebooting the PC then I will send a picture of the rest.

Froze in the lock screen.

Going to try getting into windows one more time

Now it refuses to post with GDM off....

Twice in a row

[steady lance]

No luck getting it to post again but the timings were extremely loose. Jedec loose kind of timings.

[short blade]

what voltages

[steady lance]

vSOC = 1.25
VDDP = 1.15
VDDG = Auto
VDD = 1.65v
VDDQ = 1.45v
VDDIO = 1.4v

[short blade]

you can try up to 1.3 vsoc

set vddg to 1 just in case it's autoing to something weird

[steady lance]

There is VDDG-CCD and VDDG-IOD

Set both to 1v?

Set VDDG to:
VDDG-CCD = 1v
VDDG-IOD = 1.05v
vSOC = 1.3v

[short blade]

I've always kept them the same

[steady lance]

@short blade

[steady lance]

Crashed the very second after I sent that screen shot.

I feel like this means I shouldn't bother with GDM off.

For the Secondaries, do any of them actually use formulas or just keep lowering them as much as possible?

So does tFAW below 20 work or does it get rounded up behind the scenes?

Also, any idea why my HWinfo does not show the normal CPU clocks?

It is only showing CPU Effective clocks and does not show all the different power draws.

[red terrace]

tFAW is the min cycles to do 4x tRRD commands in a row, so if your tRRDS/L are > tFAW/4 then it won't do anything bc you'll be limited by the tRRD commands

in the screenshot your tRRDS is 8, so tFAW is effectively 32

[steady lance]

Thank you for the explanation.

Is it just tRRDs for tFAW?

Because I am 8 hours stable at tRRDs 4 with tFAW 20

[red terrace]

iirc yeah, the two tRRD commands. tRRDS and tRRDL

[steady lance]

Gotcha. Wondering because the guide only mentions it for tRRDs.

[short blade]

tFAW >= 4*tRRDS, just not less

ymmv, I was stable at 4/8/16 but saw no performance difference from 8/8/32 so I just went back to 8/8/32

think 8/8/32 is the lower limit for ddr5 on paper or something, unsure if lower gets ignored or what's happening

[steady lance]

In a Buildzoid video he mentions tFAW below 32 while faster the performance gain is very low to where he said he usually doesn't bother setting it lower

@short blade @red terrace
My current secondary timings:

• tRP = 12
• tRRDS = 4
• tRRDL = 6
• tFAW 20
• tWTRS = 2
• TWTRL = 12

[short blade]

I believe many timings have formulaic relationships specified by jedec but don't have to be followed when overclocking

[red terrace]

set tFAW to 16 then

[red terrace]

faw is literally "four access window" the timing is just a limit on min cycles for those two commands

[short blade]

yeah

I mean stuff like

tWTRS = tRRDS / 2
tWTRL = tRRDL * 2
tRC = tRP + tRAS
etc

[red terrace]

okay gotcha

[short blade]

seems a lot of people on reddit assume those need to be followed when overclocking too

[red terrace]

yeah lol

[steady lance]

Gotcha so there isn't a need to follow those formulations.

[short blade]

some people have some really weird theories

like

[steady lance]

That's an answer I have been trying to figure out for ages lmao.

[short blade]

I saw someone saying tREFI must be a multiple of 256 - 1

even jedec doesn't do that lol

that one just came out of fantasyland

[steady lance]

tRFC, should it be in multiples of 32 or is that another Fantasyland answer?

Do Secondaries and tertiaries scale with VDD or VDDQ?

[red terrace]

tbh i just minimize it and then round to the next highest clean looking number for a little bit of stability lmao

[steady lance]

Hynix likely A die is what I have I think because I can run tRFC 450.

[steady lance]

I start with no post timings, loosen till post, then keep loosening till 30m stable and then if it lasts 12hrs of Karhu give or take I consider it stable enough to move on to the next set of timings. Once all of Secondaries are done I am going to run VST3, TM5 and OCCT.

[red terrace]

i go way low and no POST, then way high, then back and forth narrowing the window until i find where it works. quite efficient actually

mainly only good for when you don't know what ICs you have tho

if you have something remotely consistent then just start at normal settings for that IC and loosen/tighten from there

[steady lance]

Fair.

I used BZ's 2x32GB 6400cl26 timings as my baseline.

@short blade true or false?

"tRFC2 = Used on AM5, ensures the data integrity at high DIMM temperature, >85°C, to be confirmed how to calculate, leave on Auto.

tRFCsb = Used on AM5, to be confirmed how to calculate."

Found another source that says tRFC2 and tRFCsb are only used on Ryzen 8000 series but not on 7 or 9.

[short blade]

pretty sure they are unused

even if they are used, doesn't matter you will encounter other issues before 85C

[steady lance]

Fair lmao my RAM becomes unstable beyond 55c.

[steady lance]

@red terrace, @short blade here's my current progress.

Trying tFAW 16 then I am going to start tightening primaries.

[steady lance]

tRCDWR = 8 so far appears to be stable and Karhu speed went up 1k.

[steady lance]

It seems tRCDRD = 38 is as low as it will go for me.

[short blade]

odd primaries will only work with GDM off

so GDM on means the next step below 38 is 36

[steady lance]

I made sure all my primaries are even numbers because of that.

tRCDRD = 36 errored. So it will need to stay at 38.

I am going to tighten the other primaries some more then do a overnight run but here is my current result, I ran it through Karhu, TM5, and VST3 for an hour each.

[steady lance]

Finally I am almost done.

Just need to finish doing tRC, tRFC, then fCLK.

Testing tRC 36.

32 failed to post.

Going to see if tRC 34 posts.

Lowering tRC seems to have a noticable impact on Karhu speed.

Average test speed went up by about 1-2K and rose quicker.

[steady lance]

If my kit can run tRFC 450 does that mean it's likely Hynix A die?

Testing tRFC 400

14c ambient does wonders for temps.

[steady lance]

What is the limiting factor for running low tRFC?

I am about 1 hour in at tRFC 400 with no issues.

[steady lance]

8 hours stable in Karhu. Running VST3.

[steady lance]

Going to start raising fCLK.

[steady lance]

Any ideas on how I can get fCLK 2233 to post?

2200 fCLK seems to work fine but if I try 2233 it just refuses to post. Gets stuck in the RAM training phase.

Trying VDDGs at 1.1 and 1.15v to see if that will get fCLK 2233 to post

fCLK 2233 is giving me a headache.... I am all ears and would appreciate any advice regarding getting higher fCLK to post.

[steady lance]

Going to start working on Curve Optimizer.

@short blade what were your recommendations for testing PBO boost clock and co values?

I am currently testing with CoreCycler using Kizuna and Auto for Runtime.

Looks like Core 0 can maintain 5.7Ghz in BKT with a peak of 5.75Ghz during initialization of a test.

Apparently the 9950x3D boost clocks by default are 5.55/5.75. 5.55 is the V-Cache CCD.

Core 0T1 effective clock is staying steady between 5.78-5.8Ghz.

Core 0T2 is also hitting 5.75-5.8Ghz whenever it's under light load.

Boost clock holding steady at 5.75Ghz

Well it seems core 0 does not like -40.

Going to try -33.

Lowering CO to -33 on core 0 resulted in 5.695Ghz Fmax and 5.706Ghz Effective.

[steady lance]

I am going to try -35 on CCD0 per core and run CC to see which survive

Looks like Cores: 0, 1, 2, 5, 6 at first glance can boost to 5.75Ghz quite easily.

Hmmm I have an idea for a CC profile I want to test later for quickly checking if a core can hit the desired max frequency in very light loads.

Oh sweet, I finally leveled up.

[steady lance]

Initial CO is looking pretty good so far, @short blade.

Negative 35 CO applied per core to all of CCD0 with a +200mhz override. Every core is hitting 5.7-5.75Ghz in BKT and 5.7Ghz effective as well.

Testing using CoreCycler with the Kizuna profile set to Y-Cruncher running the following tests per core:

• BKT
• BBP
• SFT
• FFT
• N32
• N64
• HNT
• VST3
• C17

[steady lance]

@wind willow Are you still working on your CO?

CCD0 is currently doing about 5.5-5.55Ghz effective clock in Karhu Cache+FPU.

Failed VST3 so I know where to reduce.

[short blade]

corecycler never did anything for me

karhu was better then VST at finding CO instability

[steady lance]

Cache + FPU?

[steady lance]

What test suite and profile did you use?

[wind willow]

Not really. -10co and -5 co for ccd0 and 1 with med high and max cs adjustments on medium and high temp of -15 and -10.

And I can’t do mhz adjustments per core on my board, only overall. Unless theres a place other than pbo settings that idk about

[steady lance]

Gotcha.

[steady lance]

Currently I am doing CO my current values that are looking stable for CCD0 are:

-35
-40
-40
-25
-40
-40
-30
-40

I am planning on messing with Curve Shaper after getting my COs done.

[steady lance]

+200 Override for all cores which all cores are achieving.

Going to make some adjustments. It appears some of the clocks have gone down slightly as a result of a large negative co.

[steady lance]

Going to run Karhu Cache+FPU overnight then run CC during the day.

[short blade]

yes

[short blade]

default

my chip hits all core max clock in karhu so no point testing the same clock in single core in corecycler

might be different for your chip

my chip could pass corecycler all day at basically -50 all core

but in karhu core 4 errors at -20

[steady lance]

Ahh gotcha. My CCD0 does 5.75 single and 5.5 all.

[short blade]

yeah corecycler is probably useful for you then

[steady lance]

@short blade what are the long term effects of using Scaler and BCLK?

I am debating trying a BCLK overclock once I have my CO and CS settings done.

Annoyingly, Karhu failed after 9h15m.

I wish Karhu would say which core failed

[steady lance]

Nevermind, not going to mess with BCLK I forgot it changes the speed for RAM as well.

[steady lance]

@short blade if my Effective clocks is higher than my boost clock but both reach the same max clock with about 10-25mhz variance does that mean I am clock stretching?

[short blade]

sounds fine to me

[steady lance]

Average effective clock after a long run appears to be about 50mhz higher than the boost clock.

[steady lance]

I am honing in on my stable CO for CCD0.

[steady lance]

Found out one of the CO values was not stable so I decided to start from scratch doing one core at a time.

[steady lance]

@short blade @modern walrus I am trying to better understand Scaler for PBO on my 9950x3D and would really appreciate info you guys have.

From what I could gather Scaler 10x on 9000 series will allow the cores to use up to an additional 25mv so 1.3 becomes 1.325 and that it will also force the CPU to try and maintain a frequency for longer. Is there any caveats or other behavior I did not find?

[steady lance]

I am annoyed.

I have been working on Core 0 which can hit 5.75Ghz in games, browsing , downloads, etc but it is currently only able to maintain 5.7Ghz flag under BKT load in CoreCycler.

Any ideas on how I can get it to hit 5.75Ghz under consistent load?

[steady lance]

Finally got core 0 to hit 5.75Ghz in BKT test in CoreCycler. Unfortunately it is not 5.75Ghz sustained but it is progress.

Currently running -45 CO on core 0.

[steady lance]

Disabling Override and CO to run Karhu overnight for a sanity check

[short blade]

on my chip scalar 1x vs 10x has absolutely 0 difference

ymmv

AMD also says safe cpu voltage is 1.28-1.31 so might want to reevaluate running 1.325

[short blade]

this is normal

I'm playing around a bit more with CO myself right now

19/19/19/19/30/35/35 just passed 43000%

[steady lance]

Neat.

[steady lance]

Any ideas on how I could increase my sustained frequency?

[steady lance]

Welp fCLK 2200 failed after 10h6m.

Going to run another overnight test at fCLK2133

[short blade]

it's normal to not be able to hit the same boost in heavier loads as in lighter loads

[steady lance]

True, I just want to try and raise the performance.

[steady lance]

@short blade do you think running TM5 (Anta777 Extreme), Karhu 20,000% (CACHE+GPU), and 4 hours of VST3 is enough to call a RAM and fCLK OC stable?

[steady lance]

Welp my timings were not 100% stable

Ahh I just remembered I forgot to do a full test suite when I tuned the primaries

Loosening tRC from 34 back to 48 and tRAS from 50 back to 80 and rerunning the test

What does Memory Manager Not Started mean when trying to run TM5?

[steady lance]

@proven canopy my current ram tune.

RAM KIT: Kingston Fury Renegade (6000cl32-38-38-38-80 XMP) 2x32GB Hynix Dual Rank.

Ram temps max out between 48.3-52c depending on day or night.

Cooling: 180mm fan blasting directly on the ram kits from an angle to also provide the VRMs with additional cooling that pulls air from the bottom and front of the case.

Voltages:
VDDG-CCD = 1v
VDDG-IOD = 1.05v
VDDP = 1.15v

ProcODT UP is set to 30ohms in BIOS, not 240 like ZT is showing.

[steady lance]

SkatterBencher's video #85 states that setting Scaler to 10x on the 9950x3D only raises the voltage by 10mv max.

Found that and thought it might be useful knowledge to relay.

So that would mean that 10x scaler is within the safe voltage you mentioned.

[wanton yarrow]

Anyone ever come across an issue with MSI motherboards that the pc will fail to boot and will give an error message stating that the XMP profile failed and being you into the bios? Any idea what could be causing that? The ram is rated up to 6000 mhz but defaults to 4800 mhz when xmp profile is off. Maybe the voltage is too low?

[steady lance]

What board, CPU, and XMP rating? I personally have not had this issue.

[wanton yarrow]

MSI B760M-VC Wifi Motherboard, intel i5 14400f. I apologize for being new to this, where would I find the xmp rating in the bios?

[steady lance]

Should be on the sticks label.

[steady lance]

I need to get some sleep but for this issue I recommend asking in #cpus-mobo-and-memory you are more likely to receive help there.

[wanton yarrow]

Xmp 3.0

Thank you

[steady lance]

@short blade @proven canopy any idea why HWinfo refuses to display my RAM info like temps, Voltages, etc?

Tried restarting the PC, and force closing HWinfo.

[steady lance]

Seems to be fixed a few restarts later but I did notice that in the Maximum category for PMIC Undervoltage and Overvoltage, it says Dimm A2 experienced Overvoltage and Dimm B2 experienced Undervoltage during the last power down but I am not sure if this is true or just a false positive.

[steady lance]

@short blade I noticed you said you have removed the heatspreaders on DDR3, is it any different with DDR5? I was told in the OC discord to use a hairdryer to heat it up till the heatspreaders can be taken apart without much force.

[short blade]

you might not want to ask me

I basically did it with as little care as possible since they were like $2 low bin ddr3 sticks lol

funny thing is they actually turned out to be a half decent die and I did oc and bench them

they all survived even though I knocked a capacitor off of one of them

but yeah unless you want to learn how to remove heatspreaders like a complete animal, ask elsewhere

[steady lance]

Lmao

Understood.

I do have a question though.

[short blade]

do not learn from this

🤣

[steady lance]

Am I able to use Electrically Conductive Thermal Pads on the dies?

I know I can't on the PMIC but unsure about the dies.

[short blade]

i wouldn't just cause it's probably gonna squish onto the pcb

[steady lance]

Gotcha. PTM7950 then lmao?

[short blade]

maybe you can if you wanna coat the whole thing in nail polish or something

ptm7950 might not be a good idea with low mounting pressure

[steady lance]

Hmmm so I need either a thick pad or high mounting pressure.

[short blade]

i don't think you need to stress too much about the TIM for such a lower power draw component personally

6 W/mK should be enough

you can also just test out delidded first

[steady lance]

Wouldn't that be a bottleneck when the RAM is drawing 9-9.6w each?

[short blade]

delidded with active airflow is often enough of a boost from stock heatspreaders that custom heatspreaders aren't needed

my b-die was running fine delidded, i just put the funny iwiwiwi heatsinks on it for fun

[steady lance]

Hmm alright. I will try that.

I want to start getting into HWbot subs is why I am taking all this so seriously.

[short blade]

arctic mx-4 is 8.5 W/mK and people put that on 300W CPU

[steady lance]

Ignore this image, just using it to check what my stable Secondaries were.

Does tWTRS need to be tRRDS/2?

Loosened tRRDL from 6 to 8 and tWTRL from 12 to 16.

[steady lance]

@short blade may I ask you another question?

[short blade]

go for it

[steady lance]

Can running any of these Voltages cause degradation requiring me to use more SOC voltage to stabilize the same frequency/timings?

• VDDP = 1.15v
• VDDG = 1.05v
• SOC = 1.205v
• VDD = 1.65v
• VDDQ = 1.45v
• VDDIO = 1.4v

In the past 10 days or so I have had to raise my SOC by 10mv to stabilize the RAM.

Your thoughts would be much appreciated.

I am currently leaning towards me just having some setting set to low since I reverted back to my Stable settings which has looser Terts and so far it's not erroring unlike the more recent timings.

.

One more thing. I am debating redoing my RAM tune to try my hand at BCLK overclocking because I am fairly certain CCD0 and CCD1 have more headroom but I am not sure what are some things I should watch out for / be concerned about.

[steady lance]

Dang it I jinxed myself. Just errored.

The same settings that passed Karhu for 12 hours failed in under an hour now.

[proven canopy]

Update? Reset registry related settings?

If that's no help, possible your dimms don't have sensors

Buy some 6kc26 gskill since you're on am5

[steady lance]

It fixed itself after a restart.

Was really weird.

[steady lance]

Would those be better than my current kit?

[proven canopy]

Why would I recommend a worse kit?

[steady lance]

Fair enough.

[steady lance]

I need a minimum of 48GB with 64GB being ideal. Should I still opt for 6000cl26 kits?

[proven canopy]

Then get the DR AHBD kit

[steady lance]

AHBD?

Also I am going to retune my RAM.

[proven canopy]

Hynix A-die revision H

Just search amazon for "6000 c26 g.skill"

[short blade]

don't think so

[proven canopy]

There are both SR and DR SKUs https://www.amazon.com/dp/B0F1XN7DWN

[steady lance]

Oh this is awesome.

My bios calculates and shows what my expected ram frequency will be when adjusting bCLK

[steady lance]

Will do. Also may I ask why G.Skill?

[proven canopy]

If you find a better kit, go for it

[steady lance]

Gotcha. I was just curious.

Is bCLK 1.03 to high for ssds?

[proven canopy]

Most, or all? am5 boards should have a separate pcie clock. You should check.

100.3 is nothing, that's within margin of error

I think I set 105 on an sp5 board/Turin chip not long ago, nothing died with a sata ssd

[steady lance]

103 bCLK did not mean to put the period.

[proven canopy]

Unlikely, but ymmv,

More importantly, idk why you'd want to do that instead of ratio

[steady lance]

Isn't bCLK the only way other than eCLK to push CPU Clock past the +200 override without doing a static clock?

Also I can't seem to find anything regarding PCIe clock.

[proven canopy]

Impractical for daily oc

[steady lance]

That we are in agreement with but I want to try and see how far I can push my daily tune.

Then I plan to start doing some bench tunes for HWbot

[proven canopy]

Then start dropping your negative co or pushing fmax

[steady lance]

Decrease the CO to bring the increased Fmax within the voltage ceiling correct?

[steady lance]

I like 6528, lucky 8 is now in my RAM frequency

Core 0 hit 5.865Ghz!

Any idea why ZenTimings shows my SOC at 1.3v when it is set to 1.25V in bios using Override Mode?

All 8 cores on CCD0 reporting 5.85Ghz instant clock now.

Going to do CO then RAM this time around.

[steady lance]

Going to sleep and running Karhu overnight.

bCLK 102, XMP with +200 override and motherboard limits for PBO.

[steady lance]

Just got 5.8Ghz to run on all of CCD0 through multiple iterations of CoreCycler (about two hours) and no freezes.

@short blade

No clock stretching from what I can tell as all the core effectives are within 50 of instant.

[steady lance]

ST (CC BKT & BBP): 5.8Ghz
MT (Anta777Extreme) 5.5-5.7Ghz

[steady lance]

I have learned from my observations that with AM5 whatever the weakest core (needs the most CO) on the CCD is what defines the maximum V/F Curve attainable on that CCD for the highest boost clock.

[steady lance]

@proven canopy are G.Skill royals binned better than G.Skill Neos?

[proven canopy]

Royal/neo etc generally just describes the heatsink/rgb aesthetic. The XMP rating will be a good indicator of bin. The sticks I suggested above are some of the best ddr5 you can buy.

Rather, any 6000c26 gskill will be top tier for am5 bar none. Both SR and DR available.

[steady lance]

Gotcha. I will definitely get some once I get some cash in my pocket

[steady lance]

Yikes those kits cost a lot.

[proven canopy]

Because they're the best of the best. If you want good instead of great, you can pay far less.

[steady lance]

What ram kit would be better for overclocking?

1. 2x24 6000cl26 kit
2. 2x32GB 6000cl26 kit?

I am willing to get the best, just gotta save up for it

[proven canopy]

2x16

[steady lance]

I need the capacity for daily use as well.

[proven canopy]

Then 2x32

[steady lance]

Gotcha.

[steady lance]

BenchMarc, do you know where I can get my hands on 2x32 6000cl26 1.4v kits? I heard they are EOL but was wondering if some are still floating around

[steady lance]

@short blade

5.8Ghz on CCD0 is so far 4 hours stable with -25 offset to all cores on CCD0 and the cores can actually boost to 5.8 or slightly higher in light loads like BKT, BBP, and sometimes in SFT.

[steady lance]

Quick status update, did some further adjustments to my CO and 5.8Ghz now seems to be fairly stable on CCD0.
Survived
-4 hours of CoreCycler (BKT, BBP, SFT)
-1h30m Y-Cruncher (VT3)
-TestMem5 (Anta777Extreme)
-1hr Marvel Rivals (Quick Play)

[short blade]

nice

i pushed core 0 down to -24 i guess

almost at the 243 mark..

24/19/19/19/19/30/35/35

[steady lance]

I got my stable clocks a bit higher.

Went from 5.814Ghz Minimum Max clock seen in CCD0 under load to around 5.83-5.865Ghz now.

Also lowered the CO values some more. I found Curve Shaper was a really good tool to test for Fmax without because I could tell it to only affect Max clocks so if the CO is unstable then when at idle and in CC I can tell which cores are struggling much easier.

I am now 7 hours no issues on the new CO values.

• 1hr idle
• 1hr OCCT
• 5hr Karhu

Oh also out of curiosity, @short blade I put -25 co on CCD1 just to get an idea of how good those cores are. 4/8 cores did 6.02-6.04Ghz, the other 4/8 did about 5.97Ghz.

[short blade]

I bet it's heaps more fun to play with CO on the 9950x3d

fmax is so low on the 9800x3d

[steady lance]

It's a roller coaster....

The only way I was able to finally figure which core was being a pain (worst ones) was using Curve Shaper to adjust the negative CO at MAX frequency then testing for idle speeds

For example, -25 co with -8 CS sees 5.865Ghz idle speeds just fine however -25 co with -20 CS the cores struggle to go beyond 5.7Ghz idle.

[steady lance]

bCLK is worth a try?

The settings I used for the OCCT test.

[short blade]

probably not, I'm pretty much temp limited at the voltage needed to run 5425mhz anyway

would probably need to delid to push it further

can't be bothered

9950x3d is just binned much higher than 9800x3d I think

[steady lance]

I simply got lucky. No one in the OC discord group has gotten +200 stable on their 9950x3D so far (spoke with about 10 different owners).

It does boost higher by 300mhz at stock but that seems to usually be the limit.

[short blade]

fair enough

you won the lottery 3 times lol

free upgrade from 7950x3d to 9950x3d, great cores, great IMC

[modern walrus]

@frigid snow I used "By Core Usage" for 12900K OC so one core did up to 55x iirc

[steady lance]

3 times the charm quite literally lmao.

[frigid snow]

Ew now I have to be bere

[steady lance]

Had to first experience two faulty 7950x3D to get here

[frigid snow]

@fervent talon you get to stay here too

[steady lance]

Status update.

[frigid snow]

I can

[short blade]

think my 9800x3d is decent silicon lottery but my 9070xt lost hard

[steady lance]

Core 7 at -80 CO
Core 6 at -90 CO

[short blade]

💀

[modern walrus]

was 52x working for all-core? I mean I know you ran OCCT for like 30 seconds

[steady lance]

-90 CO on core 7 errored.

-80 is running fine

[frigid snow]

No it will never do a all core boost unless I have multicore enchantment on

[short blade]

i thought -50 was the limit

weird

[modern walrus]

lol I couldn't get POST with -35

[frigid snow]

Ima try rn

[modern walrus]

my board lets you do -60

what's weird is it'll let you do +100

imagine your CPU is so garbage that

[short blade]

hi gary hope you're doing well

[steady lance]

My board has no limit as far as I am aware

It lets me set -100 even

[modern walrus]

btw I saw this thing where this dude tried to work around no core multiplier for 7000 x3d by setting PBO offsets to max voltage & using BCLK

it didn't let him have more voltage tho

[frigid snow]

Hey good

I don’t think my static voltage works with TVB

[modern walrus]

how's it going? hope you're also well

[frigid snow]

It tried 5.2ghz and uhh

[steady lance]

Brb

[frigid snow]

Rip static voltage

[steady lance]

Also I think -100 might work on core 6.

[modern walrus]

ehh maybe if you disable TVB voltage optimizations

[frigid snow]

It has to go

[frigid snow]

I can always just manually set 5.2ghz and adjust voltage

[modern walrus]

but if you want to have like 1-2 cores run at like 58x or something, you don't want static

[frigid snow]

What would you do for good Undervolt

And performance

[short blade]

still job searching 💀 but otherwise good

[modern walrus]

dude looking for jobs is just horrible nowadays like

Indeed, that horrible, awful LinkedIn crap

[frigid snow]

GRRRR

GGOOD SUGGESTIONS

[short blade]

does it make an actual measurable difference? i think everything below -50 gets treated like -50 but i might be wrong

[frigid snow]

plz

[modern walrus]

it all seems like a void for data collection

[frigid snow]

Chicken nugget

[short blade]

i think -100 would be an undervolt of 0.5V? which just sounds wrong

[modern walrus]

lol remember when I tried to be slick & force core offset magnitudes further than the maximum negatives by also setting negative vCore offset

I was like wth? my clocks are higher but my R23 score is way worse

[frigid snow]

😦

[modern walrus]

that's when I learned that clock stretching existed

[frigid snow]

I’m bored

[short blade]

yeah i've been very careful to measure that performance is actually going up while tuning CO

[frigid snow]

Can I borrow 5.6ghz from one of you

[modern walrus]

set 1-core to 58; 2-core to 57; 3-core to 56, 4-core to 56, 5-core to 55, 6-core to 54, 7-core to 53, & 8-core to 52 then play with voltage offsets until it works lol

[frigid snow]

Okay bet

[modern walrus]

I ignored mine for awhile cuz figuring out 12 is such a pain in the ass

especially since I really only use 6 lol

[modern walrus]

apparently Ryzen Master will like auto OC now?

like find offset magnitude suggestions

wish I'd known this awhile ago instead of wasting time trying to find them myself lol

my Core 0 is so trash I can only do -8

it's almost like your 5800x lol

[short blade]

i don't trust that 💀

[frigid snow]

I only have 6 p cores btw

Silly good good

[frigid snow]

It didn’t do much for me

I just ended up removing the program

[modern walrus]

neither do I but might've been a decent place to start

instead of like for both CCDs then each CCD then every damn core individually

I dunno anything about it but this test seems really good at finding CO instability

even tho it's apparently for memory

I didn't even do the typical CoreCycler approach

cuz last time I did, all cores would pass

then R23 would crash & reboot

I had to make my cheatsheet again cuz when it said error on Core 15 I'm like uh

[steady lance]

It gives a general idea from what I hear. Not actual stable daily co.

@modern walrus @short blade you two think -100 CO will work on core 6?

It's currently at -90

[short blade]

check if going beyond -50 is even doing anything tbh

[steady lance]

For Max boost clock no it does not raise it unfortunately because the worst core on the CCD is what determines the max frequency for the entire CCD

.

-100 posted.

Going to run that then run -30 to do a comparison since -30 is what my confirmed stable was.

[steady lance]

@proven canopy I discovered that LN2 mode allows me to set SOC higher than 1.3v. I wanted to get your thoughts on if 1.325v SOC would be dangerous to run daily?

[proven canopy]

personally, I woulnd't care to run over 1.2 daily

I value my daily stability

[steady lance]

If it's stable is it safe?

I plan to test for stability but I want to get your thoughts on the safety.

[proven canopy]

that said, buildzoid's comments on my rock solid 3990x daily back in the day

[proven canopy]

[steady lance]

Bad wording.

[proven canopy]

To clarify, I wouldn't, nor would I recommend it. If it works for you, great, but I wouldn't risk my daily driver with such settings.

[steady lance]

I meant to say, I will be sure to test if it's stable, I am just concerned with not knowing what the safe limit is

[proven canopy]

^^^^^^^^

[steady lance]

Hmm.

Also, is 1.45v Core Voltage fine? My board defaults to 1.4V.

[proven canopy]

Run PBO + CO with negative CO for daily

[proven canopy]

As long as you do that, and your board/bios is not otherwise problematic, you're good

[steady lance]

I do understand what you mean by staying within default limits.

But I am more so curious as to how far one can push said limits without causing component failure

[steady lance]

I've read that...

I feel like there is some miscommunication.

[proven canopy]

I guess my advice as this point is to push your ship as you like until you kill it to figure out the limits

[steady lance]

I was hoping you knew what others had tried that resulted in damage so I could try to figure out at so and so voltage is when damage typically occurs

[proven canopy]

Over 1.3 is risky ish

It's not a binary condition. It's a daily, I value stability, it's another story for a bench rig.

[steady lance]

Fair enough.

I wanted to know about SOC daily limits because I think I might be able to stabilize 6700ish 1:1 with 1.32v

[red terrace]

that's a lot tbh

[steady lance]

Didn't work so I reverted back to 1.28v SOC for 6528 (could probably run lower but will do that when I am done with PBO).

[short blade]

there's a reason why agesa hard limits vsoc to 1.3

it's because chips were literally exploding/melting from boards autoing 1.35v-1.4v daily

the fact that it's named "LN2 mode" should be enough to tell you that you should probably not daily it

[steady lance]

I know but I was curious....

Also I found out why, extremely stable COs were failing me at the weirdest times. On AUTO my CPU in all instances tried to run as little as 0.5-0.7v vCORE for a all core load like VT3.

Raised CPU LLC to Mode 5 and haven't had a problem since.

Oh also, I finished tuning CCD0 and I checked to confirm, all cores on CCD0 can hit 5.8Ghz in Marvel Rivals like 80% of the time in Marvel Rivals.

Currently trying to tune CCD1 which is being very finicky.

[steady lance]

Progress! Made it to 46k in CineBench 23 Multi-Thread with just PBO and I can still squeeze more out of my PBO before doing RAM 😁

[short blade]

so it turns out my 9070xt was not a silicon lottery loser

it was just skill issue on my part

was trying to use fast timings but fast timings suck

[steady lance]

Nice.

Right now I am struggling to get high freq stable on CCD1. My goal is to get the top 10 for the CR23 MT score on HWbot.

I need to score 49,900 to do so.

[short blade]

you're going to need exotic cooling for that

also don't bother with stability testing if you're trying to compete on hwbot

none of the top scores on hwbot are even close to something you could daily

[steady lance]

For the scores I don't plan to daily.

Just submit them revert to the daily profile which is what I am currently working on.

[short blade]

ram tuning makes basically 0 impact on r23