Overclock.net banner
201 - 220 of 250 Posts

·
Registered
Joined
·
15 Posts
I updated to the AMD AM4 AGESA V2 PI 1.2.0.0 Beta BIOS(5833) for my ROG Strix x470 board but it seemed to lower scores by 2-3% and PBO for my 5800X was locked to 5025mhz boost. Where before I could clock as high as CPU would let me(5100mhz) under old BIOS(5809). Went back to old BIOS for now till they post non beta version.
That's exactly the same problem i have on my 5800x. I was trying to figure out whether it's a specific issue on Gigabyte BIOSes or maybe an AGESA 1.2.0.0 'fix' to allow >1900 IF clocks.
(i could never even post beyond 1867 on 1.1.0.0, on 1.2.0.0 i can get to 2000 although i'm not happy with the RAM and SOC voltage requirements, so i'm currently at 1966 stable)

I guess your comment does at least answer my first theory (it's not Gigabyte specific...) however i really hope it's not meant to stay like this in the final version, cause it's just extra performance down the toilet...

(edit: grammar)
 

·
Registered
Joined
·
14 Posts
That's exactly the same problem i have on my 5800x. I was trying to figure out whether it's a specific issue on Gigabyte BIOSes or maybe an AGESA 1.2.0.0 'fix' to allow >1900 IF clocks.
(i could never even post beyond 1867 on 1.1.0.0, on 1.2.0.0 i can get to 2000 although i'm not happy with the RAM and SOC voltage requirements, so i'm currently at 1966 stable)

I guess your comment does at least answer my first theory (it's not Gigabyte specific...) however i really hope it's not meant to stay like this in the final version, cause it's just extra performance down the toilet...

(edit: grammar)
I was also stuck at 1867mhz, I dropped it back to 1900mhz and tighten timings down. I'll try AGESA 1.2.0.0 when it non Beta see if there is fix. For now I'm at 5075mhz Boost and 3800mhz memory and rock stable. So I'm happy.
 

·
Premium Member
Joined
·
6,260 Posts
One last hurrah for bios 3003 before i update to a bios with AMD AM4 AGESA V2 PI 1.2.0.0 and support for Nvidia smart access memory.
Cold air benching with EK custom waterloop+TechN Zen3 waterblock :)
Curve optimizer = -30 allcore
Stable in everything i throw at it, and no WHEA errors.
View attachment 2475344
Cinebench r23 multithread = 32229 points
Cinebench r23 singlethread = 1729 points

Cinebench r20 multithread = 12441 points
Cinebench r20 singlethread = 674 points

Cinebench r15 multithread = 5404 points
Cinebench r15 multithread = 288 points

CPU-Z validator @ AMD Ryzen 9 5950X @ 4798.88 MHz - CPU-Z VALIDATOR

Some Asus realbench + Passmark performancetest numbers @ PassMark Software - Display Baseline ID# 1359214 (This machine is ranked #36 out of 156355 results globally)
View attachment 2475345

Geekbench 4 @ ASUS System Product Name - Geekbench Browser
Singlethread = 8215 points
Multithread = 74733 points

Geekbench5 @ ASUS System Product Name - Geekbench Browser
Singlethread = 1844 points
Multithread = 20054 points

Some heavy IBT high+very high and Y-Cruncher numbers:
View attachment 2475346

Did also run a full sweep of all 3dmarks, but i will post that in one other thread :)
A 5950x does much better than my 3950x with a CCX overclock of 44/43.5/43.25/43. even with memory really tight at below.

Edit: I'm really upset at the local PC store I pre-ordered a 5950x from launch day Nov. 5th. A guy that ordered the same day I did got his a month ago.

Toronto went on lockdown a week after he got it. Canada Computers isn't answering their phones, isn't answering email tickets, nada, I'm in limbo.

I AM an essential worker that does tech support at a health agency, I said that in an email to them, I AM allowed to go pick it up even with the lockdown.

And the store is still open, just their customer support is terrible. I'll never pre-order anything from them again. It'll be B&H Photo for sure.

I preordered an MSI MEG B550 Unify-X from B&H Jan. 15th and their customer support told me they are shipping it Jan. 29th, decent turnaround.


2475850


2475851
 

Attachments

·
Premium Member
Joined
·
6,260 Posts
How do you check the manufacturing date on a 5950x?

Edit: I think I figured it out, the BG number is the year and week of so mine is 2048, 2020 between
Week 48Nov. 23, 2020Nov. 29, 2020
 

·
Premium Member
Joined
·
6,260 Posts
  • Rep+
Reactions: Sam64 and 72kos

·
Registered
Joined
·
220 Posts
I don't take much stock in peeps that overclock but don't show OCCT with no WHEA errors.
Now Skeeter, he ain't hurtin' nobody.
 

·
Registered
Joined
·
31 Posts
I don't take much stock in peeps that overclock but don't show OCCT with no WHEA errors.
Shots fired! 🏹

I agree though, it's disappointing to get some great OC numbers and then realize you can't even get through a millisecond of AVX2 workload or memory pattern tests without the CPU/Memory/VRM puking up its guts.
 

·
Registered
Joined
·
147 Posts
I feel like I'm about to ask a dumb question but...there's no Curve Optimizer on the Dark Hero, is there? I see DOC switching and PBO, but there is no Curve Optimizer in the sub-menus.

Just making sure I'm not missing something.
 

·
Registered
Joined
·
15 Posts

·
Registered
Joined
·
31 Posts
Hi, how do you achieve this performance on you 5950x, it seems so far for me. Do you have some bios dump ?
Not all 5950x are created equally. I couldn't get my 5950x to break 30000 in Cinebench after DAYS of testing different settings. I made tweaks, took notes, made refinements... and wasn't getting the results some people were seeing. I eventually tried Ryzen ClockTuner which has a 'Diagnostic' feature which runs through some basic tests and gives you an idea of how good your processor is (if it falls into Gold, Silver, Bronze 'bins' for silicon quality). Mine was reported as 'Bronze'. Considering this is my second 5950x - which replaced my first 5950x that suffered from constant low-utilization/low-power shutdowns (WHEA & APIC errors), I'm happy to have a stable PC and have learned to live with my 'bronze' friend. He may not be the smartest in the school, but he's reliable!
 

·
Registered
Joined
·
99 Posts
From my experience it's all about cooling. My 5950X also didn't reach 30k in CB23 with my old 240 AiO. After i changed to a custom watercooling solution (Watercool Mo-Ra 420): 30222 CB23 points. (with Mainboard limits)

2486987
 

·
Registered
Joined
·
31 Posts
From my experience it's all about cooling.
Cooling is definitely critical, and you want to make sure that is closely monitored and being managed. If your temperatures look good and you don't see your clocks drop as the temperature goes up, then the performance cap might be somewhere else.

The "silicon lottery" is a real thing. There is a reason binning exists. It's all about finding out the optimal performance for YOUR silicon sample (given all the other variables in your system like cooling of course). Not every chip will find the same optimal point and peak performance even if every other component in the PC is the same.
 

·
Registered
Joined
·
99 Posts
True @mtavel, lottery matters. Finding out the optimal performance also takes time. Took me several weeks of testing my ram and curve values to get to this point.
 

·
Registered
Joined
·
31 Posts
Has anyone getting good Cinebench scores also run the Ryzen Clock Tuner diagnostic and know if it ranked their CPU as a Gold, Silver or Bronze sample?

Just curious. Thanks!

Here is a sample of mine:

DIAGNOSTIC RESULTS​
AMD Ryzen 9 5950X 16-Core Processor​
CPU VID: 1187​
CPU TEL: 1187​
Max temperature: 68.9°​
Energy efficient: 3.69​
Your CPU is BRONZE SAMPLE​
 

·
Old crazy guy
Joined
·
3,327 Posts
When @danakin told me he was going to send back its new 5950x cause it was failing CoreCycler even with very mild CO counts, I wasn't worried. But I should have.
It was not a good sample, it couldn't boot at FCLK 1900 and I've seen a lot of people complaining about the same issue with their new 5000.
But even his new 5950x that could run FCLK 1900 couldn't pass CoreCycler... and then I got worried. And indeed there were reasons to be.

At first I thought it was the Prime95 issue but the last CoreCycler is including the updated version.
Then I downloaded the last version and run it on my 5950x and it failed on all cores. Horror :)
It was quite some time since I've run it last time.
I've run OCCT SSE/Large/Extreme which I did run not recently but later than CoreCycler.
Failures as well on many cores... Why?

The conclusion is that AMD in AGESA 1.2.x.x changed drastically how Curve Optimizer works; at least how it behaves with mine.
But I guess it's a common issue since there are so may reporting the same problems with cores failing in OCCT and CoreCycler.
If you didn't run CoreCycler with AGESA 1.2.x.x, do it.
OCCT is more forgiving even with Large/Extreme but CoreCycler it's not.

The Scalar works differently as well, from what I remember (this may interest you @Veii).
I see the same positive 60-75mV offset on all-core but on a single thread the offset is much less 3-9mV at best.
Previously I'm pretty sure for some cores was up to 10-20mV.
I guess it's a sort of nerfing since in theory you can dial the offset with the count.
I don't like it cause made much more difficult for the worst cores to keep reliably the clock and reduced the range of the best ones.

The big change is really in how the CO counts are working.
I tried relentlessly to fine tune my counts to get the best performances but failed.
A simple base count for all the non best cores and 1 or 2 for the best ones always yielded much better results.
Which is, sadly, still true; problem is that this method doesn't work anymore.
I could get good performances and pass OCCT and CoreCycler.
Now the boosting is a bit better, very high clocks (should have suspected something was off), but it's not stable anymore.

When I last run CoreCycler, at the time was only Large, it wouldn't really keep a high clock even on the best one.
Now the best core with Huge can run at 5,134 MHz. It's massive.
But this means that all the other cores that can't run Prime at these speeds are failing miserably.
It's like they removed a layer of intelligence which was throttling the clock with very intense workloads.
Which in theory I'd say is better but it ended up badly for me with a silent air cooler :p

The counts behavior was pretty straightforward before. Now it's not.
Lower count meant, in a certain range, lower voltage and higher frequency; simple (almost) and linear.
But reducing the negative count now can have different outcomes; usually lower the voltage but not always and the frequency only jumps up when it likes the "tune".
I have the feeling that now it's going to be like finding an FM radio station with a knob.
The inter-connection between the counts still exist, when you change one count it does affect the others.

Previously changing a count by just 2 ticks on one of my best cores had a very big impact.
A much less aggressive count would make it run like one the worst cores.
Now to fix my Core 1, which is in theory my best core #1/1 (!!!), I had to go from -20 to -8.
I've lost something but much less than expected, only 50 MHz in max boost and a few points in CPU-z.
In CoreCycler still runs 5020 MHz despite -8 seems a miserable negative count.

I had to spend 10-12 hours fine tuning each core and of course the result is a sensible loss in almost all benchmarks.

Since no one cared to share how to fine tune the counts I'll share my experience.
Made some mistakes at the start and of course it costed me quite some time and loss of data.
It's very tedious and time consuming so better to avoid mistakes...

If you don't have a CO configuration start with something aggressive like -25 the best and Core0 and -31 all the rest.
Otherwise keep the current CO configuration.
Do some stability checks like running Geekbench 5 to understand if it's rebooting under load.
Adjust the counts to something that doesn't reboot.

If you have OCCT with a license start and run its CoreCycler function with SSE/Large/Extreme.
You can do a rough adjustment of the counts and take note of the Cores which are failing.
For me Core 8, 9, 1 needed an adjustment in the count to pass.
These are the first candidates to check and fix with CoreCycler.
Otherwise go straight to CoreCycler.

For CoreCycler I used dataset Huge and 6 minutes.
You should test all the datasets and also for longer but honestly... not sure it's feasible, it takes months :)
If you have the guts, try at least to run Small. I'm terrified it could crash so....

Open in CoreCycler directory the config.ini and find the coreTestOrder parameter, it's set to Default.

Below how I structured it, of course you can keep the information somewhere else.
But you need to avoid to keep open, use or run anything else running CoreCycler.
Only HWInfo with a speedy 500ms refresh and Notepad++ with the config.ini.

Code:
coreTestOrder = 1, 0,  2, 4, 13, 5, 6, 3, 10, 7, 11 12, 14, 15, 8, 9
# Core EffC VID   CO        
#    0 4981 1.315 18 < 20 4975 1.330
#    1 5021 1.386 8  < 10 5037 1.386 < 10 5027 1.390
#    2 4980 1.339 25 < 25 4943 1.323 < 25 4992 1.313
#    3 4985 1.301 22 < 22 5011 1.347 < 25 5000 1.324 < 28 4986 1.331
#    4 5108 1.391 20 < 20 5134 1.405 < 20 5132 1.402
#    5 4975 1.332 28 < 28 4982 1.322
#    6 5021 1.346 28 < 28 5046 1.360
#    7 4947 1.324 22 < 22 4959 1.329 < 25 4965 1.321 < 28 4826 1.271
#    8 4867 1.382 22
#    9 4772 1.394 10 < 10 4788 1.401 < 10 4767 1.399 < 10 4788 1.401
#   10 4885 1.384 28
#   11 4786 1.392 15 < 18 4789 1.383 < 20 4754 1.373 < 22 4816 1.374 < 25 4840 1.378 < 28 4775 1.367
#   12 4935 1.387 25 < 28 4921 1.373
#   13 4806 1.372 28 < 28 4840 1.378
#   14 4880 1.386 28
#   15 4814 1.376 28
#
# Huge 6 minutes
# vCore SVI2 1.5V
# vSOC  SVI2 1.2V
# Scalar     10
# Boost      125
# PPT        280
# TDC        165
# EDC        215
# CB23 ST 1661 MT 29326
# CB23 MT 90% VID Min-Max 1.042-1.284 EffClock 4.333-4.407-4.540
# CB23 ST 6TL Core 4 VID Max 1.369 EffClock 4.994-5.010
# CPU-Z ST 682.4 MT 12954.7 CORE4 699.2
# Geekbench5 ST 1765 MT 19038
You need to specify every core you have and put them in the order you want to test.
If you have already identified bad cores with OCCT put them first.
Fixing them will help fix others without touching their CO count.

Before fixing the counts for Core 9 and 1 testing with OCCT, the Core 15 failed under CoreCycler.
After started working properly and didn't need a change.

Then in the order I've put the Core 0 and the best ones, 2 and 4.
That was a mistake in second thought, put all the worst in reverse priority order.
Fixing them will "free" more resources for the ones you care the most.

Once you find a core failing change the order and reboot to change the count.

Let's say all good till you find an issue on Core 13:

coreTestOrder = 1, 0, 2, 4, 13, 5, 6, 3, 10, 7, 11 12, 14, 15, 8, 9

Put the Core 13 first and the Cores you already tested at the end:

coreTestOrder = 13, 5, 6, 3, 10, 7, 11 12, 14, 15, 8, 9, 1, 0, 2, 4

Now reboot, adjust the count and re-start CoreCycler.
I decided to change the count by 2 or 3 ticks depending on the Core quality and time of occurrence.
If it's a good core it's worth to try a smaller change or if it's crashing very late.
Anyway it's worth to make another round with a more fine tuning.

It's very important you keep HWInfo open and record the effective max clock and VID.
Of course it's not super accurate but it should be enough.
This info is important to verify in the future if the CO behavior changed or as a reference to restore the same if it did change.
It's also a baseline for further tuning; you know if a Core is worth to try to fine tune or not.

Record the change you made and its count.
A count change will have an impact on the other Cores.
If you see there's a difference that's worth noting, record it also if you didn't change the CO.
I didn't record the first changes and I regret it.

When you are done fixing all the cores failing, you'll have at least one record for each Core.
If you pass a cycle with all the Cores (yay), let it run for another cycle if you have time.

If you are not exhausted you can make other runs trying to improve the count for the best cores.
At last you can try to improve also the bad ones but it's less relevant.

Just as example why it's important to record the data:

Core 0 seems to under-volt with a lesser negative count (or HWInfo didn't catch it and it's the same VID), frequency almost the same.
It's a candidate to check if an even less negative count can improve.
# 0 4981 1.315 18 < 20 4975 1.330

Core 1 needed an adjustment not to fail after other Cores got fixed.
Not sure it can be improved as it was originally at -20. Also VID and clock didn't change much from -10 to -8.
Seems to be at the limit already.
# 1 5021 1.386 8 < 10 5037 1.386 < 10 5027 1.390

Core 2 was influenced by the changes, fine tuning could recover the small loss and maybe improve it
# 2 4980 1.339 25 < 25 4943 1.323 < 25 4992 1.313

For Core 3 I know already I can't do much:
# 3 4985 1.301 22 < 22 5011 1.347 < 25 5000 1.324 < 28 4986 1.331

But maybe I can improve a bit Core 4 now and regain the small loss:
# 4 5108 1.391 20 < 20 5134 1.405 < 20 5132 1.402

At the bottom I've added some values and benchmarks as a reference to compare future changes.
Don't ask me about the VSOC for some reason got locked at 1.2V and I didn't notice...
Hope it doesn't make a difference when I change it back :p

Update: and it does; VSOC 1.16V at LLC3 can't pass, VSOC 1.175V LLC2 did it.

The loss in boost from the fine tuning was substantial:

2489603


To this:

2489604


About 400 points in CB23 MT, 5 points in CPU-z ST and 80 in MT.
In Geekbench 5 the MT loss was about 100-200 points.

But I hope to recover some with more fine tuning.
 

·
Registered
Joined
·
104 Posts
Hi guys my 5950x was instable: random rebooting in game and it was impossible to complete a geekbench run while every other bench/stress test was 100% stable. I just found out that this was caused by the load line calibration setted to 4, how could i set this value (or adjust other voltages) to keep my overclock stability? I was running dynamic oc switch with ccd1 4.7 and ccd2 at 4.6 with the pbo curve at all core -25, voltage offset +0.025
 

·
Old crazy guy
Joined
·
3,327 Posts
Hi guys my 5950x was instable: random rebooting in game and it was impossible to complete a geekbench run while every other bench/stress test was 100% stable. I just found out that this was caused by the load line calibration setted to 4, how could i set this value (or adjust other voltages) to keep my overclock stability? I was running dynamic oc switch with ccd1 4.7 and ccd2 at 4.6 with the pbo curve at all core -25, voltage offset +0.025
Very likely some of your cores are not stable.

Use CoreCycler to verify, -25 all core is extreme.

High LLC and PWM are required to sustain low CO counts.
You can also set Boost clock from 0 to 50 to reduce the frequency boost, performances will be slightly impacted.
Unless you have a royal cooling the cores will not keep more than 5100 MHz on sustained load.
 
201 - 220 of 250 Posts
Top