AMD Curve Optimizer any guides / experience

[freeagent]

im really tempted to try your config.
which cooler you got ?

I am using Thermalright FC140.. I think the only place to get it is Aliexpress but I could be wrong

 

[Zach_01]

I'm gunna guess 5600x get all the cherry picked cores, you lucky undervolting bastards

Usually though, low segmented SKUs are getting the worst silicon. At least that was the case with R5 3600. If you think about it, there is no (reasonable) point for AMD to cherry pick silicon for the “low end” part of the whole series.
If any CPU “needs” the best cores that would be the 5800X.

I’m guessing that since 5000 series the 7nm node has improved much overall. 5600X being the only single CCD 6core(active) CPU on 8core CCD, maybe this is giving it the edge.

Just a thought I’m not going to pretend that I know facts.

 

[Mussels]

Usually though, low segmented SKUs are getting the worst silicon. At least that was the case with R5 3600. If you think about it, there is no (reasonable) point for AMD to cherry pick silicon for the “low end” part of the whole series.
If any CPU “needs” the best cores that would be the 5800X.

I’m guessing that since 5000 series the 7nm node has improved much overall. 5600X being the only single CCD 6core(active) CPU on 8core CCD, maybe this is giving it the edge.

Just a thought I’m not going to pretend that I know facts.

well, with the 5600x and 5800x starting as the same chip, the odds on the 5600x having all better cores is actually higher - only needs one weak core to not make the 5800x cut, leaving 6-7 good ones as a possibility

 

[freeagent]

Just throwing this out there, but I have a very hard time getting my 5600X stable at more than 4850.. 4900 is good for some light benching. So looking at how cores alone scale I don’t think it’s that good of a bin..

 

[Taraquin]

I'm gunna guess 5600x get all the cherry picked cores, you lucky undervolting bastards

It's a bit easier to UV 5600X due to lower SC speed and fewer cores Freeagents sample is unusual though. Getting a 5800X with 2 extra cores and 200MHz higher stock speed stable at -30 CO is a bit trickier

 

[freeagent]

I don’t know.. I didn’t know anything about AM4 when I bought into it.. just noob tunes that’s all.. someone should send me their cpu to play with to see if I can get it to do the same thing

Don’t worry you would get it back

 

[Zach_01]

well, with the 5600x and 5800x starting as the same chip, the odds on the 5600x having all better cores is actually higher - only needs one weak core to not make the 5800x cut, leaving 6-7 good ones as a possibility

Comparing those two I could agree, yes.

It's a bit easier to UV 5600X due to lower SC speed and fewer cores Freeagents sample is unusual though. Getting a 5800X with 2 extra cores and 200MHz higher stock speed stable at -30 CO is a bit trickier

5600X(s) are also 8core parts (CCDs). Have 2 cores disabled that didn't make it. Defective from beginning or after through evaluation or just disabled because AMD wanted 6core CPUs.

Actually you have to think every 8core CCD as exactly the same on the 7nm node. From 5600X to 5950X.

5950X gets the best CDDs and 5600X the worst(?). AMD has to segment all core performance equivalent to core count as much as possible.

Does the 5950X (2x8core) performs twice (+100%) as a 5800X (1x8core)? No...
5950X performs about +55~85% better than a 5800X but on about the same power consumption (140~142W). No doubt 5950X has the better cores and already a "high" negative V/F curve compared to 5800X by AMD.

Now, comparing the same way the 5900X (2x8core but 2x6c active) with 5600X (1x8core but 1x6c active) is a little different because of their different power consumption. 5900X has 87% higher power draw from 5600X. Does it perform 87% better? Is it by avg higher or lower than 87%?
Well if you check benchmarks you'll see that a 5900X performs by avg around 90% better than 5600X. So the cores of 5900X are same or even a little better than 5600X.

So far
5950X better than 5800X
5900X slightly better than 5600X

5900X lands in between the 5950X and 5800X in terms of quality (best cores) because all 3 have the same power draw 140~142W.

So far
1. 5950X
2. 5900X
3. 5800X

Where does 5600X goes though? Easy...

5900X has 50% more cores, and around 30~40% more performance, but the same power compared to 5800X. You can say its a big difference in quality.
5950X has 33% more cores, and around 25~30% more performance, but the same power compared to 5900X. You can say its a substantial difference in quality.
So if 5600X has cores with same or slightly worst quality from 5900X then we have it all.

In terms of core quality the order is this:
1. 5950X
2. 5900X
3. 5600X
4. 5800X

Its not hard to understand that 5600X(s) have the best headroom for negative V/F curve from stock settings because of their low stock power draw compared to all other 5000 SKUs.
Its product segmentation.

 

[blu3dragon]

Early on the rumor was that 5800x was unavailable and priced relatively close to the 5900x because it needed a completely working die and AMD could equally use those for a 5950x.
5900x max stock boost clock is 4950MHz I think, vs 4850MHz for the 5800x. 5800x needs 8 cores that good, while 5900x only needs 6 per die. It depends on the yield AMD is getting. Are there more chips with 6 good cores that can make 4950MHz, or more with all 8 cores working that can't quite make the 4950MHz cut? My guess would be the former.

Either way, I think it is reasonable to say that a 5800x will have a better die than a 5600x. In the worst case, you could disable the two worst cores on a 5800x to make higher clock speeds :-D (Although I don't know a way to pick which actual cores are turned off)

The other factor here (as mentioned already) is that a CO of e.g. -5 on a 5800x is not equivalent to a CO of -5 on a 5600x, since the curve set from the factory is not the same. As I understand it, the curve is actually set per cpu, so even from one 5800x to another, -5 does not mean the core will be getting the same voltage. -5 on two cores on the same chip doesn't even mean that.

 

[Taraquin]

With OCCT v9, how do you setup for testing CO stability?

 

[Nordic]

We just need an automated way to tune in the curve optimiser and we'll allllll be happy.

I just started messing with core optimizer and isn't this the truth. It would be really nice if a script could change the core optimizer value and stress test from the bios. It would only require simple if then logic.
1) Set Core X to -5
2) Stress test Core X for Y minutes
3) If the test completed successfully, set Core X an additional -5, and repeat step 2. If the test failed, set Core X +1, and repeat Step 2. Do not test the same value twice.

Blue Dragons testing script is helpful but not the whole package. It seems like it would be trivially easy to run some sort of simple stress test at the bios level and let the motherboard figure this out automatically. Or in the OS if we could access the curve optimizer settings.


I do have a question. Is it worth trying to tune curve optimizer if I am on air cooling and thermally limited?

 

[tabascosauz]

I just started messing with core optimizer and isn't this the truth. It would be really nice if a script could change the core optimizer value and stress test from the bios. It would only require simple if then logic.
1) Set Core X to -5
2) Stress test Core X for Y minutes
3) If the test completed successfully, set Core X an additional -5, and repeat step 2. If the test failed, set Core X +1, and repeat Step 2. Do not test the same value twice.

Blue Dragons testing script is helpful but not the whole package. It seems like it would be trivially easy to run some sort of simple stress test at the bios level and let the motherboard figure this out automatically. Or in the OS if we could access the curve optimizer settings.

I do have a question. Is it worth trying to tune curve optimizer if I am on air cooling and thermally limited?

Technically AMD probably has access to CO through Ryzen Master since it's all under the same submenu, but we all know that it's not the most polished/reliable software around, frankly wouldn't trust it even if it did...

People entrusted their Renoir APUs to 1usmus CTR and it ran 1.5V all core through them to verify an "overclock" soooo I would just stick to BIOS and forget about this software convenience BS.

If you want to be reasonably stable then the default corecycler settings script is fine, so 6 minutes per core. Run through it 2-3 times and you're good, all in a day's work.

But I still stick to the 68 minute All FFT config, insane amount of time to test but at least I know it's 110% stable. Been running my 5900X curve and it still holds after switching to the B550 Unify-X. Still haven't proven it wrong.

Then there's also the OCCT test that does basically the same thing but is a bit better structured/polished. I don't use it yet but @freeagent likes it

Might explore some different testing methodologies for my 5600G. Currently at -15 no problems but fine tuning still needed.

 

[freeagent]

I use OCCT, Linpack Xtreme, TM5, Superpi 32m, y-cruncher, and a few others..

If I were to run core cycler my system would fail in seconds lol.

So its absolutely stable, but its not..

 

[Nordic]

I am running core cycler 1 minutes per core starting from a value of negative 10. If it passes the test, I increase the value by 1. If it fails I decrease it by 1. When I find the highest value a core can use, I have core cycler ignore that core.

Usually the cores that fail, fail immediately. This method is painfully slow. It would be really nice if I could automate it. So far after 5 runs I have only found the final value for Core 9. Oddly enough, the cores I thought were my best are the ones that keep failing.

After I have completed this process, I should have a fairly good idea of what my curve looks like. I intend to let core cycler run for an hour per core and use other stress tests to ensure stability after.

I am unclear on a few things though. I don't know if I am going to have to go through this process again when I install water cooling. I am thermally limited right now. I was also planning on trying out x2 scaling to see if that makes a difference, but that isn't worth trying until I have water cooling anyway.

 

[tabascosauz]

Oddly enough, the cores I thought were my best are the ones that keep failing.

Welcome to 2CCD haha, sample size still too small but I suspect average 5900X/5950X don't usually have much real undervolt headroom on their 2 preferred cores

Mine looks like this:



For preliminary testing I'd say default 6min is worth it, given that it cycles every few seconds. Anything stable on default config already shouldn't give you major problems in daily usage, just give it a couple loops as stability may not be consistent. Then if you REALLY want to ensure stability do the All FFT hour or long OCCT testing, but I'm just OCD

 

[Nordic]

Cores 0,1, and 4 are down to CO -7 and going down. Core 9 is settled at CO -11. The rest of the cores up to -16 and just keep going. It would be really nice if this was automated. Not sure how far this will go.

It seemed stable at all cores -10 too. I was considering leaving it there because it worked fine and a 15% multi threaded improvement is not bad at all. Good thing I am stability testing.

 

[tabascosauz]

Cores 0,1, and 4 are down to CO -7 and going down. Core 9 is settled at CO -11. The rest of the cores up to -16 and just keep going. It would be really nice if this was automated. Not sure how far this will go.

It seemed stable at all cores -10 too. I was considering leaving it there because it worked fine and a 15% multi threaded improvement is not bad at all. Good thing I am stability testing.

On the 6min config I was "stable" at -7 and -10 on 0 and 1. No real instability symptoms from any of the cores so clearly usable, honestly, but once I went to the exhaustive 68min config it was quickly apparent that those two cores could go no further than -2 and -7.

But yeah, unless you get really lucky and can push more than -15 on those preferred cores only, little ST benefit since the algorithm cannot be forced to use any other cores. As in the picture Core 2 easily becomes the best core after -30 CO, but I cannot ever use it. MT gets a big boost but might still be hotter at iso-voltage

Optimumtech said that CO benefits ST temps greatly, but he was testing at -15 to -30. Zero difference here at -2 and -7.

 

[freeagent]

I should take the time to play with it like that.. its just so tedious lol..

 

[Nordic]

I should take the time to play with it like that.. its just so tedious lol..

Extremely tedious. I have a timer going so that I know when to come back, see what failed, input new values into the bios, and start all over. In the mean time I have made tremendous progress on a yard project. I am very happy that I have Core Cycler to automate some of the steps.

 

[freeagent]

Honestly.. I just used SuperPi 32M to tune my single core boost, every other program I use will just bring all the cores down to one speed and float around by a few MHz, that's why I didn't try to exploit each core because I didn't think windows was smart enough to use the cores like that.

It was not that easy to do lol.. those guys have some skill for sure

SuperPi - 32M with BenchMate overclocking records @ HWBOT

 

[Nordic]

It took several hours of the smallest amount of stability testing but I finally have rough numbers.

Core 0	-7​
Core 1	1​
Core 2	-30​
Core 3	-18​
Core 4	-5​
Core 5	-30​
Core 6	-30​
Core 7	-30​
Core 8	-30​
Core 9	-11​
Core 10	-28​
Core 11	-24​
Core 12	-30​
Core 13	-30​
Core 14	-18​
Core 15	-30​

 

[Taraquin]

What is the quickest way to test CO stability? Get a rough indication? Considering trying to get +200 pbo stable. Everything at -30 allcore and +50 is rock stable since May, but I get rare reboots if I try +100 or more. +200 was stable for 24H, then suddenly pc restartet at medium load.

 

[tabascosauz]

It took several hours of the smallest amount of stability testing but I finally have rough numbers.

Default config test? That's some commitment right there took me the better part of a month to work mine out and I was still cutting corners by sticking to multiples of 5 on ten of the cores

Positive offset is some kind of mystical beast on these chips lol, this is probably the 2nd time I've ever seen it

What is the quickest way to test CO stability? Get a rough indication? Considering trying to get +200 pbo stable. Everything at -30 allcore and +50 is rock stable since May, but I get rare reboots if I try +100 or more. +200 was stable for 24H, then suddenly pc restartet at medium load.

Corecycler script, don't change the config from default, close all possible background apps and disconnect from internet. Run the script, it tests each core for 6 minutes then cycles to the next one.

Let it keep going and observe every once in a while until it's gone through every core 2-3 times. Get thru 3 iterations without errors and I'd say that's good enough for most ppl. Shouldn't take long at all on a 6-core.

I'll probably be doing the same tonight for my 5600G, sitting on lazy -15 all core right now but I need to test.

 

[Taraquin]

Default config test? That's some commitment right there took me the better part of a month to work mine out and I was still cutting corners by sticking to multiples of 5 on ten of the cores

Positive offset is some kind of mystical beast on these chips lol, this is probably the 2nd time I've ever seen it



Corecycler script, don't change the config from default, close all possible background apps and disconnect from internet. Run the script, it tests each core for 6 minutes then cycles to the next one.

Let it keep going and observe every once in a while until it's gone through every core 2-3 times. Get thru 3 iterations without errors and I'd say that's good enough for most ppl. Shouldn't take long at all on a 6-core.

I'll probably be doing the same tonight for my 5600G, sitting on lazy -15 all core right now but I need to test.

Went through 2 full cycles now, all at -30, except one core at -28, had error in 1 cycle at one core, but -28 fixed it. Satisfied with performance

 

[freeagent]

With the stock clock range I get errors @ -30 CO on my 5900. After that I just didn't care too much. I don't think windows uses the cores the way core cycler does, not even remotely..

 

[Nordic]

Default config test? That's some commitment right there took me the better part of a month to work mine out and I was still cutting corners by sticking to multiples of 5 on ten of the cores

Positive offset is some kind of mystical beast on these chips lol, this is probably the 2nd time I've ever seen it

As per my earlier post, I ran core cycler at 1 minute per core until it didn't fail. When it did fail, it almost always failed within seconds. This gave me the very rough numbers I shared above. While I slept last night, I ran Core Cycler 20 minutes per core and it only found an error on Core 0. I intend to Run Core Cycler until I get no errors anymore then try OCCT.

To my knowledge, I am the third invididual in this long thread who had a positive offset. I suspect I needed a positive offset because I have LLC set to 3.