AMD Curve Optimizer any guides / experience

[tabascosauz]

Yea, extra profiles, different sizes, not sure what to choose and why Up until now I always chose the smallest size when running p95 manually, but now I'm not sure after what you told me in relationship with the Huge preser.

I left everything on default and set it to only cycle through one core. If it passes 12 hours, I'd move onto the next. I guess that does enough iterations.

Still use P95 Small as a torture test for Intel OCs. It's just not the right tool for the job here. Like if you were stressing memory controllers, you would run P95 Large FFT.

If you're still doing 6 minute config and doing 1 core at a time, just run it a couple times and go to the next. Save the long tests for a comprehensive config. Huge FFT takes what, 30-45 min to actually test the entire range? 6 minutes only gets around to testing a small snippet, it's a quick and dirty test.

I would still test at least 2 cores at a time. That way they alternate and can take a break from the heat. Preferably cores on different CCDs as well, unfortunately not possible for 5800X

 

[Octopuss]

What exactly does an iteration mean in the context of this script btw? Like one pass of a specific FFS size?

It looks like the defaults might be a little misconfigured though, if default size is Huge and time 6 mins...

 

[tabascosauz]

What exactly does an iteration mean in the context of this script btw? Like one pass of a specific FFS size?

It looks like the defaults might be a little misconfigured though, if default size is Huge and time 6 mins...

1 iteration is just running through the length of the test once on 1 core. You can set the # of rounds in config. It'll run in the core order specified in config, then repeat.

It isn't 100%, but 6 min Huge is a good balance of speed and effectiveness. Strictly speaking it's probably not "stable", as you'll quickly find if you run eg. 1 hour All afterwards, but if you can pass 6 min Huge you *shouldn't* really have any big issues day-to-day with gaming and regular use. It's a bit different than mem stability where no number of errors is acceptable, because chances are you won't be able to match corecycler clocks in daily use (unless your CPU is so good it easily tops out at +200MHz)

Properly testing is best, but it takes a long time. Better for 6-8 core parts, but it took me like 2.5 weeks to run the entire CPU through 1h15m All.

 

[Octopuss]

I am still pretty confused by all this. It looked dead simple at first...
How did it take 2,5 weeks though? If I take a look at the config and add up the numbers there, even if I double them, that's a few hours at most.

I only had one coffee today though!
Maybe I am missing something essential.
The script is pretty complicated.

 

[tabascosauz]

I am still pretty confused by all this. It looked dead simple at first...
How did it take 2,5 weeks though? If I take a look at the config and add up the numbers there, even if I double them, that's a few hours at most.

I only had one coffee today though!
Maybe I am missing something essential.
The script is pretty complicated.

4-5 iterations per core, 1h15m per iteration, 2 cores testing per night. Some cores required re-adjusting so more nights retesting new settings. And even then 10 of the cores are set in multiples of 5 (-10, -15, -20 etc) because I just couldn't be bothered at that point.

With default config, you can easily be done in just 1 day. Just do that, and if you find that it's not stable enough, then go for the long tests.

I never really had any real issues with settings that pass default config, on either 5900X, 5600G or 5700G. Just wanted to be doubly sure for my 5900X daily.

 

[Octopuss]

So anyway, if I want to test it really thoroughly, several hours on each core with Huge size (or maybe All) should be pretty safe?

 

[DoLlyBirD]

timings and aida bench attached, any advice on my timings?

 

[tabascosauz]

So anyway, if I want to test it really thoroughly, several hours on each core with Huge size (or maybe All) should be pretty safe?

Should be fine yeah if you have the time for it.

timings and aida bench attached, any advice on my timings?

This is the Curve Optimizer thread, nothing to do with timings or AIDA performance. There is an AIDA thread, Ryzen owners club and some others I think.

Share your AIDA 64 cache and memory benchmark here | TechPowerUp Forums

 

[DoLlyBirD]

Should be fine yeah if you have the time for it.



This is the Curve Optimizer thread, nothing to do with timings or AIDA performance.

So my posts on the previous page regarding TSME you replied to shouldn't be here either?

 

[tabascosauz]

So my posts on the previous page regarding TSME you replied to shouldn't be here either?

I won't delete those, that started off with CPU performance and CPU-Z results and eventually meandered into the other stuff. Ryzen discussion tends to become a bit of a melting pot anyway. But let's try to keep purely mem-related stuff in the other threads.

 

[DoLlyBirD]

I won't delete those, that started off with CPU performance and CPU-Z results and eventually meandered into the other stuff. Ryzen discussion tends to become a bit of a melting pot anyway. But let's try to keep purely mem-related stuff in the other threads.

My point was, I started off asking about curve optimiser and PBO settings and then it got on to memory timings and YOU started the TSME talk that lasted for nearly a page... so now I should just start a new thread and disregard the previous posts even though there are literallly tons of the same?

Go to BIOS > AMD CBS and disable TSME. You are penalizing yourself by 7-10ns

Boost Clock Override tops out at 200 and has been for a long time

TSME aside, latency is high for 4000CL16 but Im assuming that's because you haven't touched tRFC yet

Use zentimings if you want to get into mem stuff

Would bios settings for AMD overclock/CBS etc

I can't find any TSME setting in my bios at all, I knocked the TRFC down a tad as it was about 560 using default settings, though wanted to be careful as have had issues lowering it too far too soon in the past with these sticks, though that was on a Ryzen 1600AF when I could only run max 3400mhz stable

tRFC gets kinda easier to run "tighter" ns-wise the higher speed you go. Because you need more VDIMM to push lower tRFC, but each kit has a hard lower limit to tRFC (in number not ns) where it won't boot anymore.

If it doesn't do sub 200ns tRFC, it's not B-die. No matter how horrible the bin. Even legacy E-die does 160ns. 140ns can be a little tough for bad B-die but 160ns should be no prob

Some vendors now put TSME on the main page under CBS. For others it should still be under UMC settings or something under CBS. And then under Security I think. MSI should be fine, don't recall it being that hard to find on my Unify-X.

View attachment 244473

I have MSI B450m Pro-A Max, it's a turd of a board.. still can't see any resemblance of that setting..

Interesting thing though with the TRFC, tried to run low 300's on my 1600AF and it didn't like it at all though that was likely due to the 1600AF IMC and not the RAM, I'm at 400 now, I shall go deeper down the rabbit hole lol

View attachment 244474

b-die is the way to go for lowest tRFC,
low tRFC works very well on amd, every -10trfc nets you a good .1ns better latency

here's my patriot 4400cl19 kits
View attachment 244475

Yes, but at the same time AIDA is also disproportionately a sucker for tRFC, doesn't quite translate to real performance. Big jumps on tRFC make big gains, but smaller differences below 140ns is not quite as noticeable. Other timings have real impact as well but AIDA ignores them.

Tradeoff between VDIMM and tRFC. 1.56V set on Gigabyte boards is closer to 1.6V real, since all Gigabyte boards overvolt VDIMM between 0.03-0.06V.

View attachment 244477



I'll take a look at my Unify-X BIOS later tonight and send some screenies of CBS menu.

Can't help you much if you don't provide your zentimings. I'm confident you can do 145ns or better, but you should expect to bump VDIMM for it. tRFC isn't free.

ZenTimings

ZenTimings is a simple and lightweight app for monitoring memory timings on Ryzen platform.

zentimings.protonrom.com

Ya, cba with that, I'll figure it out myself, thanks.

 

[GoldenX]

Tested both the new Ryzen Master and HYDRA-PRO, both are unstable. Good think I did it manually.

 

[tabascosauz]

My point was, I started off asking about curve optimiser and PBO settings and then it got on to memory timings and YOU started the TSME talk that lasted for nearly a page... so now I should just start a new thread and disregard the previous posts even though there are literallly tons of the same?

Ya, cba with that, I'll figure it out myself, thanks.

I didn't reprimand you, just saying that we have a number of threads that will generate more discussion for pure mem stuff. You are welcome to create a new thread if you don't feel any existing ones are appropriate. We haven't done a great job of keeping things strictly on the original topic here, and I did have a hand in that. But that was also two weeks ago, and you expressed an initial interest in tweaking CO. On topic has to start somewhere.

 

[Ibizadr]

timings and aida bench attached, any advice on my timings?

First advice, focus only one thing and the other after first concluded with success. Put default bios on it and focus on ram first. Try 1t GDM off, 3800/1900 mclk==fclk. You need to test ram a lot. And after that come to this thread to improve your cpu, but next time discuss on Aida thread.

 

[DoLlyBirD]

First advice, focus only one thing and the other after first concluded with success. Put default bios on it and focus on ram first. Try 1t GDM off, 3800/1900 mclk==fclk. You need to test ram a lot. And after that come to this thread to improve your cpu, but next time discuss on Aida thread.

I've used default XMP 3000 settings, bear in mind I'm running at 3800 now with the same basic settings, have tweaked trfc and tfaw used curve optimiser to boost to 4450mhz with 3800mhz RAM for 1:1

 

[Ibizadr]

I've used default XMP 3000 settings, bear in mind I'm running at 3800 now with the same basic settings, have tweaked trfc and tfaw used curve optimiser to boost to 4450mhz with 3800mhz RAM for 1:1

Your timings needs improvements take a look in Aida topic

 

[Octopuss]

Is it possible to be able to use higher negative offsets a year after thoroughly testing a CPU? I don't think BIOS updates make much of a difference.
On the other hand, I was running Prime95 manually with the smallest or small FFTs back then. Also ran two threads of it on the core I waz testing, and I'm not sure what the CoreCycler script actually does, because I don't quite understand that from the config file. Logic says you should run two threads with SMT/HT on, but...

 

[AVATARAT]

Is it possible to be able to use higher negative offsets a year after thoroughly testing a CPU? I don't think BIOS updates make much of a difference.
On the other hand, I was running Prime95 manually with the smallest or small FFTs back then. Also ran two threads of it on the core I waz testing, and I'm not sure what the CoreCycler script actually does, because I don't quite understand that from the config file. Logic says you should run two threads with SMT/HT on, but...

I wrote here how I test it.

 

[Octopuss]

Also wtf, I increased the negative offset on two cores and now my all core (I think, it's what Afterburner shows) boosts are lower? That makes no sense. Previously I usually saw 4800MHz and now it fluctuates between 4650 and 4725.

I wrote here how I test it.

How would the script ovevolt cores? That's nonsense.

 

[tabascosauz]

I wrote here how I test it.

Overvolting cores? PBO On causing Vcore to spike over 1.5V is common but CO causing it is a new idea.

Also wtf, I increased the negative offset on two cores and now my all core (I think, it's what Afterburner shows) boosts are lower? That makes no sense. Previously I usually saw 4800MHz and now it fluctuates between 4650 and 4725.


How would the script ovevolt cores? That's nonsense.

If you want to properly gauge your clocks in any sort of sustained load, use HWInfo. Either look only at the Effective Clock metric, or enable Snapshot Polling in settings (right click tray icon) and look at Core Clock/Effective Clock. Nothing else (except maybe Ryzen Master) properly shows a realistic clock. Hitting 5150/5250MHz "core clock" (just an example) without having similar Effective numbers to back it up means it will never sustain long enough to actually make a difference in benchmarks or effective clock.

Real 4.8GHz all-core is very, very hot on Ryzen. Afterburner OSD is a classic culprit of bamboozling you into thinking the CPU is working a lot harder than it actually is.

 

[AVATARAT]

Also wtf, I increased the negative offset on two cores and now my all core (I think, it's what Afterburner shows) boosts are lower? That makes no sense. Previously I usually saw 4800MHz and now it fluctuates between 4650 and 4725.


How would the script ovevolt cores? That's nonsense.

The problem is not in the script, but if you use it, it will throw errors until you put enough or more voltage in the Curve on the core that fails. So you just need to test it.
With SSE the cores work on a higher frequency and need not so high voltage to not fail, and that is what I use with OCCT.

 

[Octopuss]

I have just realized I never touched the PBO limit.
Does increasing it by the usual +200MHz have any negative implications?
What exactly does it do anyway?
And what is the default frequency cap of 5800X?

 

[oobymach]

I have just realized I never touched the PBO limit.
Does increasing it by the usual +200MHz have any negative implications?
What exactly does it do anyway?
And what is the default frequency cap of 5800X?

Stability in CoreCycler is not guaranteed with +200 but regular p95 will be stable. It pushes every core to perform an additional 200mhz faster so if you were getting 4.65ghz you'll get 4.85ghz. Also you may have different results in curve with the extra speed you may need more voltage.

 

[tabascosauz]

I have just realized I never touched the PBO limit.
Does increasing it by the usual +200MHz have any negative implications?
What exactly does it do anyway?
And what is the default frequency cap of 5800X?

+200 raises the global frequency limit. Stock for 5800X is 4850MHz, so it goes up to 5050.

Global limit is only a theoretical ceiling. The problem is it's really silicon dependent. You will notice in HWinfo the global limit goes up to 5050MHz, but only works if the cores usually picked for ST tasks (usually top 2 ranked cores) are actually capable of sustaining more than 4850MHz effective. eg. on a mediocre CPU you can change Boost Override all you want, but it won't make a difference.

Also, stable undervolt at 4850MHz may no longer be stable with higher boost override. You'll just have to test and see. If you are already maxing out at 4850MHz effective, good chance you can go further.

 

[AVATARAT]

I have just realized I never touched the PBO limit.
Does increasing it by the usual +200MHz have any negative implications?
What exactly does it do anyway?
And what is the default frequency cap of 5800X?

You will increase performance with this 200MHz but you need to play again with CO to stabilize it (add more voltage to be stable).