Overclock.net banner
7961 - 7980 of 12770 Posts
Joined
·
144 Posts
Well in my 51.2ns entry i clearly write its done in safemode under the "proof" tab.

And i did just get a 0.2ns difference between safemode and my everyday bloated windows.

Would say very few people takes the time to make a "bench-only" OS install, and for all others safemode is the best option.. If Veii cant cherry-pick the numbers in each tab by running them one-by-one that is just tough luck. (yes i know i have attitude about this, but that is because a certain someone was campaign for my results to get removed from the spreadsheet)
I was wondering why since you're only at 1900 when there's plenty of WHEA filled bullshit on the sheet, but maybe there should be an asterisk + note or something on your best result where it says 5950x since you're running only 1 CCD. I know it's in the image but people don't see that right away when they're just quickly sorting and comparing.

Otherwise idk what the problem is, you may as well have just slotted in a temporary 5800x if you had one that's basically all you did
 

·
Registered
Joined
·
327 Posts
I was wondering why since you're only at 1900 when there's plenty of WHEA filled bullshit on the sheet, but maybe there should be an asterisk + note or something on your best result where it says 5950x since you're running only 1 CCD. I know it's in the image but people don't see that right away when they're just quickly sorting and comparing.

Otherwise idk what the problem is, you may as well have just slotted in a temporary 5800x if you had one that's basically all you did
Dont think i can change my entry on that document after the owner have "locked it".
My dual rank flat 14 is still going strong after one hour :)
2489929

This is how i tested and binned my 4 memory sticks in the different channels:
1-2-3-4
2-1-4-3
3-1-4-2
4-2-1-3
1-3-2-4
2-1-3-4
It took some time to test all these combinations, but in the end the "1-3-2-4" setup was the best layout which allowed me to run these new timings

*edit
failed after 1hour and 15 min :(
single error 2

*edit2
quick memtest in dram calc is easy
Computer Personal computer Font Screenshot Parallel
Back to 25 cycle TM 1umus..
 

·
Premium Member
Joined
·
6,366 Posts
The CL16 4000 almost $100 cheaper and only 1.4v vs 1.5v for the CL14 3800, so the 4000 might be higher binned I'm thinking.

Edit: Both are b-die.


G.SKILL Trident Z Neo Series 32GB (2 x 16GB) 288-Pin DDR4 SDRAM DDR4 4000 (PC4 32000) Intel XMP 2.0 Desktop Memory Model F4-4000C16D-32GTZNA
  • DDR4 4000 (PC4 32000)
  • Timing 16-16-16-36
  • CAS Latency 16
  • Voltage 1.40V


or


G.SKILL Trident Z Neo Series 32GB (2 x 16GB) 288-Pin DDR4 SDRAM DDR4 3800 (PC4 30400) Desktop Memory Model F4-3800C14D-32GTZN

  • DDR4 3800 (PC4 30400)
  • Timing 14-16-16-36
  • CAS Latency 14
  • Voltage 1.50V
 
  • Rep+
Reactions: TimeDrapery

·
Old crazy guy
Joined
·
3,519 Posts
Now I'm getting somewhere... For me, the errors appeared to resolve when I adjusted my RTTs so I'm imagining this is a powering issue pertaining primarily to resistances
I'm slowly trying to make a 2T profile that runs almost the same as the 1T with setup timings now.

But I'm wasting most of my time calibrating the CO counts right now...
Found out Scalar above 6x with this new AGESA has an impact on how negative you can go with the count, exactly 2 counts.
But at 10x with higher counts is a bit faster; I'm still in the middle of the testing session :p
 

·
Premium Member
Joined
·
6,366 Posts
I'm slowly trying to make a 2T profile that runs almost the same as the 1T with setup timings now.

But I'm wasting most of my time calibrating the CO counts right now...
Found out Scalar above 6x with this new AGESA has an impact on how negative you can go with the count, exactly 2 counts.
But at 10x with higher counts is a bit faster; I'm still in the middle of the testing session :p
Edit: I did better in multicore R23 with 150 boost vs 200 boost, running Scaler 10, by 100 points too. :)
 
  • Rep+
Reactions: TimeDrapery

·
Old crazy guy
Joined
·
3,519 Posts
Edit: I did better in multicore R23 with 150 boost vs 200 boost, running Scaler 10, by 100 points too. :)
Since my cooling is a bit undersized, I use 125 MHz boost.
Works better but I have to check with the new CO counts.

Maybe you could test as well if with Scalar 6x you can drop a couple of counts more on CO.
I got more or less same results with CB R23 and Geekbench5 but 150 points more in CB R21.
Not sure yet how reliable are the scores though.
My best core in CPU-z scores 702 with Scalar 10x and 699 with Scalar 6x and 2 counts less.
It's really head to head, have to run more benchmarks.

In the meantime I've ordered the Ripjaws, should come before Saturday in theory.
 

·
Registered
B550 AORUS MASTER, 5800X, 16GB (2×8GB) g.Skill DDR4-3200CL14 memory, XFX RX 5500 XT
Joined
·
171 Posts
I'm slowly trying to make a 2T profile that runs almost the same as the 1T with setup timings now.

But I'm wasting most of my time calibrating the CO counts right now...
Found out Scalar above 6x with this new AGESA has an impact on how negative you can go with the count, exactly 2 counts.
But at 10x with higher counts is a bit faster; I'm still in the middle of the testing session :p
Righteous! What are you seeing performance-wise, relative to your 1T set with setup timings, so far?
2489933

As you can see, my latest efforts produced another lone... ERRRRRORRRRR #4 😂😂😂😂😂

Oh well, try, try, try again...
2489934

2489935

That's very interesting about the CO/scalar behavior changing up with the latest and greatest AGESA release... What do you think that says about how the FIT/SMU has changed with this AGESA? How's throttling compared to prior versions?
 

·
Old crazy guy
Joined
·
3,519 Posts
Righteous! What are you seeing performance-wise, relative to your 1T set with setup timings, so far?
2489936


Need a bit more write bandwidth to feel in peace. Not that is really that useful but...
Latency is extremely stable and RTT mild.
I want to check if I can go a bit lower with tRFC with same or higher VDIMM.
Or if I can keep it as it is and go lower.

2489937



As you can see, my latest efforts produced another lone... ERRRRRORRRRR #4
It's becoming a regular :D

That's very interesting about the CO/scalar behavior changing up with the latest and greatest AGESA release... What do you think that says about how the FIT/SMU has changed with this AGESA? How's throttling compared to prior versions?
I'd say there's less throttling which makes it better if you have good cooling.
But also less throttling on its own and hence more difficult if you don't.
It's not really about FIT/SMU but more about the how PBO boosts with single threaded workload.
 

·
Premium Member
Joined
·
6,366 Posts
Since my cooling is a bit undersized, I use 125 MHz boost.
Works better but I have to check with the new CO counts.

Maybe you could test as well if with Scalar 6x you can drop a couple of counts more on CO.
I got more or less same results with CB R23 and Geekbench5 but 150 points more in CB R21.
Not sure yet how reliable are the scores though.
My best core in CPU-z scores 702 with Scalar 10x and 699 with Scalar 6x and 2 counts less.
It's really head to head, have to run more benchmarks.

In the meantime I've ordered the Ripjaws, should come before Saturday in theory.
Oh, I'm still on A210 BIOS, I found the A24 hurt me performance-wise and I like all the unlocked options on A210.
 
  • Rep+
Reactions: TimeDrapery

·
Old crazy guy
Joined
·
3,519 Posts
Oh, I'm still on A210 BIOS, I found the A24 hurt me performance-wise and I like all the unlocked options on A210.
Same for me, stayed on the A22 for a while but it doesn't help with USB randomness.
And I missed too much the XOC options.
 

·
Registered
Joined
·
3,105 Posts
I'm slowly trying to make a 2T profile that runs almost the same as the 1T with setup timings now.

But I'm wasting most of my time calibrating the CO counts right now...
Found out Scalar above 6x with this new AGESA has an impact on how negative you can go with the count, exactly 2 counts.
But at 10x with higher counts is a bit faster; I'm still in the middle of the testing session :p
Scaler just shifts what voltage values upwards along the vfcurve during boost. If you don't have good cooling, 10x will end up hurting you because of the increased heat. There is also the increase in power consumption to deal with when doing that as well, so you're pbo budget will be eatin' into.

The different co values people are seeing with different scaler values most likely has to do with the voltages pbo is using depending on what the scalar value is. What inevitably ends up happening is with the lower scaler values, you aren't reaching the higher v/f points on the curve, so you can be more aggressive with the CO value (more negative) since it is much easier to undervolt the lower the voltage you go (which is exactly how the CO curve works, for this reason).

With the higher scalar values, you will have in increase CO values (less negative) as you are now using higher voltage bins which will be way more sensitive to undervolting.

It also explains why some may be able to use higher fmax settings with lower scalar values as pbo won't be reaching to those boost clocks nearly as much except for quick, lightweight loads.
 

·
Registered
Joined
·
327 Posts
Wish me luck :)
Dual CCD running 4x memory sticks at 1900fclock T1 gdm disabled with flat CL14 timings + tRFC 240
2489940

Never had that high memory read numbers @ 1900:3800..
Going to bed now, wonder how many red errors i will see after 25 cycle tomorrow =7
 

·
Registered
B550 AORUS MASTER, 5800X, 16GB (2×8GB) g.Skill DDR4-3200CL14 memory, XFX RX 5500 XT
Joined
·
171 Posts
View attachment 2489936

Need a bit more write bandwidth to feel in peace. Not that is really that useful but...
Latency is extremely stable and RTT mild.
I want to check if I can go a bit lower with tRFC with same or higher VDIMM.
Or if I can keep it as it is and go lower.

View attachment 2489937




It's becoming a regular :D



I'd say there's less throttling which makes it better if you have good cooling.
But also less throttling on its own and hence more difficult if you don't.
It's not really about FIT/SMU but more about the how PBO boosts with single threaded workload.
Dude, the performance you get from your setup is awesome... Not too much longer until I'll go to Zen 3 now, I'm really looking forward to the switch! I'm also looking forward to see further results from you!

It is becoming a regular, this go-round I've made it past the duration tested where it appeared during the last iteration... Happy enough with that so long as I didn't simply delay its arrival 😂😂😂😂😂
2489945

Aha, I see... It's good to hear that AGESA development is actively bring driven forward by AMD rather than languishing as it seemed to in times past (even though there was plenty to work on/with)!
 

·
Registered
Overclocking - latest technology
Joined
·
358 Posts
The CL16 4000 almost $100 cheaper and only 1.4v vs 1.5v for the CL14 3800, so the 4000 might be higher binned I'm thinking.

Edit: Both are b-die.


G.SKILL Trident Z Neo Series 32GB (2 x 16GB) 288-Pin DDR4 SDRAM DDR4 4000 (PC4 32000) Intel XMP 2.0 Desktop Memory Model F4-4000C16D-32GTZNA
  • DDR4 4000 (PC4 32000)
  • Timing 16-16-16-36
  • CAS Latency 16
  • Voltage 1.40V


or


G.SKILL Trident Z Neo Series 32GB (2 x 16GB) 288-Pin DDR4 SDRAM DDR4 3800 (PC4 30400) Desktop Memory Model F4-3800C14D-32GTZN

  • DDR4 3800 (PC4 30400)
  • Timing 14-16-16-36
  • CAS Latency 14
  • Voltage 1.50V
Hello @KedarWolf, the second model is the one I have and they are working very well with your times. I'm trying to get 1T, when I could, I got a significant increase in bandwidth and lower latency, but at some point some BSOD came along. But I still haven't tried the "magic" CadBus Setup values 56-56-56? Have you tried it?

Why would the 4000 be better binned? For the performance I saw here, I don't recommend this model for its price/performance . I would choose a 3600CL14 and save a few bucks or even the same model 3800CL14 but single rank with a lot of headroom for OC


Sent from my iPhone using Tapatalk Pro
 

·
Registered
Overclocking - latest technology
Joined
·
358 Posts
Well, the "Vodoo" thing works :) @ManniX-ITA plase tell us how to reach to this setup, I never heard or read about it before.. But now, I'm working with 1T without Wheas nor BSOD.. I will continue testing.. but, pretty amazing. There's some place to read about this CADBUS Setup values? I didn't find anything googling it. what is the rational??

I will match tCWL to TCL so I can lower one CK tRDWR and see if read write performance increase..
2489961
 
Joined
·
144 Posts
So I think these are my final 1.450v 4x8GB 3733 53.5ns settings, tRCDRD 14 could literally last 3 hours apparently before failing so just bumping up to 1.460v could probably be enough to stabilize it but I don't really want to.
2489971
2489972

It's always going to bug me that I can't boot 1900 FCLK but unless I buy a new 5800x or motherboard I think it's impossible and I don't know if it's possible to get rid of the 1933+ WHEA either. I don't even know if anyone else with a MSI B550 Tomahawk has 1900 FCLK, I'm sure someone does but it's like nobody owns this motherboard but me, no others with 1900 working to try to copy, assuming they found some magic setting or combination to get it working on this motherboard/BIOS. I feel like I just haven't found the secret but maybe there actually isn't one guess I'll never know
 

·
Registered
Joined
·
273 Posts
Interesting.... after I updated to 1202 I seem to get 600-630mbps in speed tests, downgrading and I'm back to 510 again.
 

·
Registered
Joined
·
2,117 Posts
i'm working on it and trying to stabilize :)
dual CCD 5950x with 4x8gigs sticks

tRCDRD 14 was only possible after i binned my memory sticks..

View attachment 2489922
View attachment 2489924
View attachment 2489925
Yup, as mentioned before, to be in with a chance to run 4 x 8GB with [email protected] you need to be fortunate enough to have 4 "good" sticks otherwise you need to go start binning sticks to find the combination that can do it!

Wish you luck in getting there, ive found that if you can get past the first few cycles without throwing an error 0, 10 you have a good chance of getting them stable, the difficulty is finding the right combination of setting to power the DIMMs correctly, simply increasing voltage in this scenario often proves counter productive.

Those settings in your later posts sure look familiar

:D :D

Interesting.... after I updated to 1202 I seem to get 600-630mbps in speed tests, downgrading and I'm back to 510 again.
Interesting ....

:)

@ManniX-ITA
How come youve dropped tRDRDSCL/tWRWRSCL to 2 while running 32GB ??

Have you found a way to stop the drop in read bandwidth when compared to 4/5 rather then 2/3 ??
 
  • Rep+
Reactions: domdtxdissar

·
Registered
Joined
·
24 Posts
It's always going to bug me that I can't boot 1900 FCLK but unless I buy a new 5800x or motherboard I think it's impossible and I don't know if it's possible to get rid of the 1933+ WHEA either. I don't even know if anyone else with a MSI B550 Tomahawk has 1900 FCLK, I'm sure someone does but it's like nobody owns this motherboard but me, no others with 1900 working to try to copy, assuming they found some magic setting or combination to get it working on this motherboard/BIOS. I feel like I just haven't found the secret but maybe there actually isn't one guess I'll never know
It isn't your motherboard. There are plenty of us that can not POST at FCLK 1900, but have no problem POSTing at FCLK 1933, 1966, 2000 etc. Obviously anything 1933+ will spam the event viewer with WHEAs.

I have tried everything, the board (Dark Hero+5900x) will not even try to POST at 1900. 1.4vSOC, 1.3 VDD/VDDG, tried low voltages, LN2 mode. Doesn't matter, it simply refuses to even try.

It is just one of the perks of owning AMD hardware, you get one random surprise bug that will stick with you until you upgrade back to Intel. I should've stuck with my ivy bridge until Intel came out with something decent.
 

·
Overclock the World
Joined
·
3,136 Posts
@Veii
View attachment 2489896

forgot to enable BGS, what is it for ?
vsoc is at flat 1.100 when TM5 is running.
why does the MB say vdimm is at 1,52x volt, when i enter 1,5v ?

im pretty happy with that result, but again, where can i go from here ?
14s with the same trfc ? maybe voltage was not enough before.
im a little afraid when i enter 1,51 volts (thats why its at 1,50), since the number turns red than... 🙃
best regards and thanks for help everybody!


edit: for some reason zentimings shows BGS as disabled even when i enable it (?)

regards,

pete
Good job, it's 0.2 down - and you prepared the dimms to run on GDM off
But i am sorry to tell you ~ that TM5's result is only 15-18% worth something
It shows that you don't have a catastrophic failure and likely games would be fine with it
But it doesn't show that you won't thermal throttle, and it doesn't show that predicted tREFI + tRFC range is correct

Knowing you used tRFC mini (well i did for you)
Means, tRFC won't fail after a time
But there is a difference between techical theory and real world results

Soo please either edit the MT.cfg config (inside the /Bin folder) to 25 cycles
Or grab this old reuploaded one, which has 20 cycles preconfigured

In order to actually proof thermal stability, we need to be sure the dimms remain stable @ their highest thermal peak
Which means, either running furmark while TM5 runs (soo the GPU heats it up)
Or letting a TM5 test run beyond 1h
Around 45min is required for dimms to reach Thermal Equilibrium, and around 6 cycles are required to notice and tRFC1 failure
19 cycles are required to notice any tREFI and tCKE failure ~ powerdown and a whole refresh cycle.

Soo because memory can and did often fail the last test of cycle 19, the absolute minimum is 20 cycles
Which is 1:30h for 16gb , or 3h for 32gb sets
Please retest :)
Also when using Aida64
Put yourself to High-Performance powerplan, push display poweroff to 2h, system hibernation to disabled & disable USB suspension (advanced powerplan settings, above minimum % powerstate)
This will prevent random DPC calls while TM5 runs and while Aida64 runs
Another advice,
Run Aida64 as soon as possible while the OS loads ,then let it finish
By the time it finishes, the OS will have loaded all their useless services & you can run another one (before TM5 !)
Doing so (after the first Aida64 run which loads it's own Kernel Drivers) ~ it will allow the CPU to start using it's powerstates
Soo the CPU will hibernate fully (if enabled) and the test results will be consistent and peaking up to CPUs capability & settings

If you take the first load, the thread sheduler will pioritize Microsoft "delayed" services and your result will be wrong (L3 especially)
If you run Aida64 after TM5 or with ZenTimings, the ZT kernel and both ZT & TM5's random refresh will do random DPC calls
This eats around 400-600MB/s bandwidth, slows cache down and increases tested latency
Soo to make this as correct and consistent as possible:
  • run Aida64 asap on OS start
  • wait a bit , like 10sec ~ the test takes long enough soo useless services will be suspended by the thread scheduler as Aida64's kernel enforces Realtime mode
  • after the first realtime load, the CPU will forcefully suspend itself (if supported) soo what you test is idle to high perf lantency
Last point might appear slower , but it's as close to reality as possible. The rest are peak boost tests and do not match to normal day-to-day usage. Also can randomly track microsoft service triggers and give a slower or unrealistic unstable result :)
OC'ing T-Force 4133 cl18
Post with TM5 20cycles , or edit it yourself
btw, what IC's asus use in proarts' VRM implementation ?
It's really "bad" ones from my personal viewpoint.
12+2 Vishay SiC639 , 50A
on a ASP1106-PWM-Controller (rebranded RT3667BE) which only supports 4+2 phase mode

They cheaped out on the PWM controller and had too tripple it up
Soo it's in reality:
4 Phase vCore * 3
1 Phase SOC, one DRAM
They should've used a Renesas one which at least supports 7+1 phases to make this a dual 6 phase board, instead a 4 phase one
VRMs according to 5800X test heat up to 80c ~ but the board has full switching frequency control & phase control + loadline modeling

Potentially i am less inteligent than the ASUS engineers, and their 4 phase option has better ripple ~ but somehow i doubt this
Anywho , i think it's fine ~ but i'd be worried with a 5950X on it.
210A on 100% usage is kiinda peaking the specs-sheet (160A vcore + 50A SOC)
A 5900X stock should be fine, a 5800X stock should also be just borderline fine (5800X OCs itself stronger than a 5900X)
But we'll see how it behaves on my take & if fMAX override does anything to the 5600X :)
If it's WHEA hell, then i'll just put a 1200AF or 1700X on it ~ although i'd be worried with a 180A 1700X, haha
(eh it was fine with a B350 Toma peaking 105c ~ this shouldn't be thaat bad 😅)
EDIT:
As we know, peaking the specs sheet doesn't help Ripple at all
a doubled 6 phase potentially would be better (with the option to switch it to full phase , which we can with ASUS's ext volt controller)
~ well not using Vishay 50A stages at all and going rather with 70A or 90A Renesas Mosfets on a 300$ board, would be even better
But hey, i'm neither EVGA engineer, nor have ASUS connections to judge them for this ~ maaybe it's fine. I'll hold back critique till i can verify it myself :)

A B350 Tomahawk with a mini 55mm blowimatron, was fine with over 200A load ~ this should not be even close as catastrophic as my 1700X set
Yet the B350 Toma sustained my 3467MT/s C14-14 set ~ soo thermal peaking VRMs don't always have to result in unsable ripple 😋
I'll give it a chance and see how bad it really is, have definately worked with worse things
 
7961 - 7980 of 12770 Posts
Top