Trying to understand Ryzen 3000 series boost speed variations

[Zach_01]

I had read that a while back, but rereading it now make it interesting reading. In R20 my single core results do not match there, with PBO on or off - they hit 520 whilst my very best so far is 502, but my multi core is 100 points higher, also for myself PBO does seem to increase multi core scores.
I feel that PBO is such a finicky thing even between the same setup but at different times, that it looks to be that there are no 'rules' or guidelines that can be given. Yes you Might increase your performance by a percent or two, but nothing significant beyond besting your best benchmark by a few points, certainly nothing usable in real use. By far the biggest factor that many didn't appreciate and still don't is temperature. i did realise that and the reason why I went for the best air cooling that I could at the time (even going to the expense of adding another Noctua Chromax fan to the NH DI-15S). I know that by far the best method is to get an open cooled loop and I am saving up for that, but at £400 roughly and limited funds, that won't happen next week either.

Yes temperature has big impact on clocks for ZEN2. Even if the throttle temp(95C) is never reached, or even getting close.
When I first build the system (August, ~33C ambient, CPU max ~75C) I was studying HWINFO values during various usage scenarios and I never saw “CPU package power” more than ~75W for all core load and ~45W for single core.
Now with 22 ambient, CPU max 62~63C the max CPU package power is hitting 88W for all core and 55W for single. Max PPT allowed is 88W for my CPU.

Yes Intel will, but also remember that AMD would have also been developing and fine-tuning their own 7nm products that would be already out. So, as to whether they match each other or one remains ahead only time will tell.

I have a distinct feeling that given the designs of both Intel and AMD, that Intel will likely always be ahead in the fastest speeds and that AMD will be ahead in multicore usage. Unless of course either or both of them radically change the designs of their CPUs.

I feel that the clock wall being near is relative though. It is near, but I also do not expect it to be reached within the next couple of years. I am sure that the actual production process will be far more of a hindrance as the chiplets become thinner and that will then slow the progress to the thinnest wafer that can be made.

It may even be that we see a radical change in design. For instance, instead of going thinner and thinner, reaching an obvious end point, what if say AMD were able to develop and expand upon the Infinity Fabric to increase the distance between cores without loss of communication speed? That way you even see the size of the CPU increase significantly to allow distance between everything and dissipate the generated heat better, thus increasing performance that way?

Intel will eventually switch architecture and may loose any high clock advantage in the future. Intel’s current tech was designed for monolithic low core count and high clocks dies. I believe this will come to en end soon (2021-22).

Manufacturers already know how things are going for the next decade or a little more. 7nm -> 5nm -> 3nm -> end of current UV tech or unknown factors. They also want the higher density (within reason) for optimal space management and for higher die count production per wafer. Higher yields is also a key factor for them to make as much profit they can get and offer a reasonable priced products at the same time. Chiplets(small dies) are helping with this. So they will be forced to use less voltage and clocks to achieve that density and will hunt performance gains through IPC and other optimizations. I will find a video about all this and post it here.

There it is... published March 2019.
It’s very interesting and I suggest to watch it all.

 

[tabascosauz]

No idea What, or even Where, FIT voltage limiter is, or even if it is configurable ?

I have to ask all this because the MB manual doesn't even have a PBO section in the BIOS chapter!

When I run/ran P95 i hyst started the auto configured stress test, which i believe has a mixture of everything in it?

I guess that I will have to go into BIOS to see exactly what mine does?
Or would HWINFO show me the max settings for PPT/EDC/TDC ?

FIT is set by AMD and is not configurable. It's like a knock sensor for your car's engine, which knows exactly where the limit of the engine is to prevent damage through knocking/detonation (here, unsafe voltage). But unlike the car's ECU, FIT can't be changed by the user like the ECU can be tuned.

Like before and like I've said, you'll find the PBO menu under AMD Overclocking in the Peripherals tab.

P95 gives you an option of Smallest, Small, Large and Blend when you open it. Use either Smallest or Small for CPU stability (whichever gets hotter on your system) and Large for memory stability. Blend isn't great because P95's main selling point is that it's extremely intensive and generates a lot of heat; that makes it useful to see where the equilibrium of your cooling setup lies. When the temps basically stop climbing and settle around a particular point, you know how much your system can take. Blend would simply shift the load into something else and allow the part that's heated up to cool down again.

The 1usmus plan is easily explained. The earlier your Windows version was when you switched to 1usmus, combined with the farther your best cores are from Core 0, the more you stand to gain from the 1usmus plan. If your chip has its best cores already in the first CCX (like yours), you stand to gain relatively little; Windows is using them by default. If your chip has its best cores in the last goddamn CCX and the first CCX is home to hot garbage of an excuse for silicon (like mine), there can be increased gains for those who have been shafted by AMD and TSMC; that's still subject to if those best cores are actually able to clock appreciably higher than the shitty ones. It's the exact same story for 1909 vs 1903, you may or may not notice a difference. The ridicule towards 1usmus really only shows some reviewers need to check their privilege, having received relatively well-off chips either in binning or core quality and distribution, and get off their fucking high horse.

 

[Waldorf]

@tabascosauz
not looking into difference in perf etc, just looking at scheduling/which cores are being used, that plan changes nothing.
i read a test from a website (dont remember name, but one i knew for years so i know it wasnt a 12y old in the grandparents basement),
as well as some guy not being happy that his stuff wasn't running on the best clocking cores, both times they could show,
that there was no change which "cores were picked".

it does improve things for some because of they having the wrong "settings" (win etc) or un-tweaked power profiles,
but not for most of us running all the latest (bios/drivers), and so far most that looked under the hood,
stated its done by adjusting sleep of cores/cxx (delay/depth/wake up).

seeing that i couldn't measure any negative impact compared to the latest amd profiles that came with the chipset update,
i decide to use that (further tweaked).

but hey, if it works better on yours (vs the amd ones), why not.

 

[tabascosauz]

@tabascosauz
not looking into difference in perf etc, just looking at scheduling/which cores are being used, that plan changes nothing.
i read a test from a website (dont remember name, but one i knew for years so i know it wasnt a 12y old in the grandparents basement),
as well as some guy not being happy that his stuff wasn't running on the best clocking cores, both times they could show,
that there was no change which "cores were picked".

it does improve things for some because of they having the wrong "settings" (win etc) or un-tweaked power profiles,
but not for most of us running all the latest (bios/drivers), and so far most that looked under the hood,
stated its done by adjusting sleep of cores/cxx (delay/depth/wake up).

seeing that i couldn't measure any negative impact compared to the latest amd profiles that came with the chipset update,
i decide to use that (further tweaked).

but hey, if it works better on yours (vs the amd ones), why not.

What are you trying to say?

I've read a few of the "debunking" tests. Did any of them test beyond the Ryzen 3000 samples they already had on hand? Did any of them account for the improved scheduler in 1903/1909 that they would be running at the time of testing? Did they level the playing field by eradicating variation in AGESA versions? One hell of a "trusted" reviewer. That's like me dismissing hundreds of other F150 owners suffering oil consumption on the 5.0 Gen3 motor just because mine happened to be problem free. Conduct a thorough test with a more significant sample size, and now we're getting somewhere with results that the community can actually make use of.

I've literally just described the conditions under which you might stand to experience some gains: without the Windows scheduler improvements that render the plan anywhere from slightly to largely redundant, with a chip that has suboptimal distribution of its best cores, with a chip that actually has an appreciable delta (read: thus, actual potential gain) in performance/clocks/voltages amongst its cores. And most definitely, without running AGESA 1.0.0.4-based BIOSes with actual CPPC controls that didn't exist at the time of the plan's release.

 

[lorry]

FIT is set by AMD and is not configurable. It's like a knock sensor for your car's engine, which knows exactly where the limit of the engine is to prevent damage through knocking/detonation (here, unsafe voltage). But unlike the car's ECU, FIT can't be changed by the user like the ECU can be tuned.

Like before and like I've said, you'll find the PBO menu under AMD Overclocking in the Peripherals tab.

P95 gives you an option of Smallest, Small, Large and Blend when you open it. Use either Smallest or Small for CPU stability (whichever gets hotter on your system) and Large for memory stability. Blend isn't great because P95's main selling point is that it's extremely intensive and generates a lot of heat; that makes it useful to see where the equilibrium of your cooling setup lies. When the temps basically stop climbing and settle around a particular point, you know how much your system can take. Blend would simply shift the load into something else and allow the part that's heated up to cool down again.

The 1usmus plan is easily explained. The earlier your Windows version was when you switched to 1usmus, combined with the farther your best cores are from Core 0, the more you stand to gain from the 1usmus plan. If your chip has its best cores already in the first CCX (like yours), you stand to gain relatively little; Windows is using them by default. If your chip has its best cores in the last goddamn CCX and the first CCX is home to hot garbage of an excuse for silicon (like mine), there can be increased gains for those who have been shafted by AMD and TSMC; that's still subject to if those best cores are actually able to clock appreciably higher than the shitty ones. It's the exact same story for 1909 vs 1903, you may or may not notice a difference. The ridicule towards 1usmus really only shows some reviewers need to check their privilege, having received relatively well-off chips either in binning or core quality and distribution, and get off their fucking high horse.

Yep, I know where PBO is, it's currently set to advanced and the +200MHz option selected. I couldn't remember seeing a FIT setting though, that's why I asked
Thanks re prime95, so small for CPU and large for memory, got it.
I'm pretty sure that by the time I installed 1usmus I was already on 1909 (the Ryzen power plans were already an option and I Had selected those).
I'd like to see what say Gamers Nexus reports re 1usmus, as I tend to trust his testing (states the How they test and the Why), but from their reports and videos etc they tend to think that there isn't any big improvements to be made. Even so if there's a few percent to be had for free, it would be daft not to say thank you very much and grab it!
To be honest I doubt that I will see much in the way of any realistic improvements in my CPU, AMD have already got it set up at close to max performance already, sure a few tweaks here and there but nothing earth shattering. I have a feeling that for myself the only major improvements were in the ram, and even there it is likely close to what can be got out of it without Much better knowledge than I currently have.

 

[Zach_01]

FIT is set by AMD and is not configurable. It's like a knock sensor for your car's engine, which knows exactly where the limit of the engine is to prevent damage through knocking/detonation (here, unsafe voltage). But unlike the car's ECU, FIT can't be changed by the user like the ECU can be tuned.

I tried to find some info about this FIT and I couldnt. And I never heard/read about such term either.
Can you link us somewhere we can read about it? I'm interested in such things, and I believe @lorry too...

 

[tabascosauz]

I tried to find some info about this FIT and I couldnt. And I never heard/read about such term either.
Can you link us somewhere we can read about it? I'm interested in such things, and I believe @lorry too...

References to FIT are really patchy, as with most things AMD (try looking up resources on the SVI2 bus for Ryzen 3000 specifically, kek). A Reddit comment here, a AMD forum post there...some do point to an OCN post by the Stilt, "Strictly technical: Matisse (not really)". A chunk of the info there is deprecated due to it being written after release and subsequent AGESA having changed how the CPUs, work, but the references to FIT are in the "The overclocking capabilities" section.

More importantly, there's a lot under there that affirms what I've been saying about Zen 2, that in order to attain the advertised clockspeeds, the chips have to go beyond what is considered efficient for N7FF, which lies around the area of 3700X and lower SKUs' base clock: 3.5-3.6GHz. After this, necessary Vcore rises like magma up through a volcano. It explains why shitty chips have to be made into 3600/3700X, and why the lucky "good" 3600 can be OCed to rival 3700X/3800X speeds. At or below 3.6GHz, Vcore is low but not amazingly so (~1.0V), and power consumption and thermals are nearly dormant. If AMD can just stop trying to bite off more than they can chew, EUV 7nm could be a revolutionary efficiency leap for 4000.

Interestingly, he notes that Matisse has nothing like an AVX-512 offset that Intel has, which keeps thermals in check in those AVX scenarios. From what I've been seeing lately in P95 Small specifically as opposed to the normal clocks in Smallest, I can't help but wonder if this is still true.

 

[lorry]

Strictly technical: Matisse (Not really)

07/08/2019 6:33 PM (GMT) - Update on the bios issue on Crosshair VIII Hero motherboard ("the thing"). Earlier today I received a response to my inquiries from ASUS. The response was rather technical and I cannot go into the specifics of what exactly it involved. However, it confirmed my...

www.overclock.net

CCX Overclocking Tool for AMD CPUs Updated With New Features

Today an overclocking utility for AMD Ryzen CPUs called "Work Tool" has been updated with new features. The tool enables per-CCX overclocking, which is said to enable additional overclocking performance if one CCX is more capable than the other, so the whole CPU doesn't have to run at the speed...

www.techpowerup.com

A comment posted by 1usmus in that thread -

"Delete this application and forget about existence, you will kill your processors!
This is hacking FIT limits, the last line of protection for the processor "

 

[mtcn77]

If you are into PBO overclocking/undervolting, I suggest using EDC for undervolting purposes and PPT for overclocking intentions. I find the standard power-performance curve being most easily adjustable single-handedly by that respective order.

 

[lorry]

If you are into PBO overclocking/undervolting, I suggest using EDC for undervolting purposes and PPT for overclocking intentions. I find the standard power-performance curve being most easily adjustable single-handedly by that respective order.

Do you have any links you could suggest where I could read up some more on it?

 

[mtcn77]

Do you have any links you could suggest where I could read up some more on it?

Just wrote a review on it: [Based on a Reddit post]

 

[tabascosauz]

Strictly technical: Matisse (Not really)

07/08/2019 6:33 PM (GMT) - Update on the bios issue on Crosshair VIII Hero motherboard ("the thing"). Earlier today I received a response to my inquiries from ASUS. The response was rather technical and I cannot go into the specifics of what exactly it involved. However, it confirmed my...

www.overclock.net

CCX Overclocking Tool for AMD CPUs Updated With New Features

Today an overclocking utility for AMD Ryzen CPUs called "Work Tool" has been updated with new features. The tool enables per-CCX overclocking, which is said to enable additional overclocking performance if one CCX is more capable than the other, so the whole CPU doesn't have to run at the speed...

www.techpowerup.com

Seems like you just have to set manual voltage to stop it from going to 1.405V Vcore. Not sure how your BIOS reacts; mine straight up doesn't POST if I start splitting up clocks by CCX, so I'm not too enthusiastic about this tool. Also, I can't seem to find any reference to it beyond its initial release in July, which isn't promising, considering all the changes that have happened since then.

If you are into PBO overclocking/undervolting, I suggest using EDC for undervolting purposes and PPT for overclocking intentions. I find the standard power-performance curve being most easily adjustable single-handedly by that respective order.

Didn't seem to do anything positive at those settings. I still bench best and perform best day to day in CB R20 with PBO off and -0.075V offset. Running stock PPT and TDC with slightly lower EDC and stock Vcore just behaved like a worse maxed-out PBO experience. If undervolting gives you gains, there's a much easier way and surefire way to do it than messing around in PBO.

 

[mtcn77]

Seems like you just have to set manual voltage to stop it from going to 1.405V Vcore. Not sure how your BIOS reacts; mine straight up doesn't POST if I start splitting up clocks by CCX, so I'm not too enthusiastic about this tool. Also, I can't seem to find any reference to it beyond its initial release in July, which isn't promising, considering all the changes that have happened since then.



Didn't seem to do anything positive at those settings. I still bench best and perform best day to day in CB R20 with PBO off and -0.075V offset. Running stock PPT and TDC with slightly lower EDC and stock Vcore just behaved like a worse maxed-out PBO experience. If undervolting gives you gains, there's a much easier way and surefire way to do it than messing around in PBO.

Well, if there is any, please provide directions as to how. This normally entails a larger debate as to whether PPT or EDC provides the higher benefit to which I myself base my verdict on, so if you would have just told to undervolt I would thusly state no, PPT is better and even further is EDC. PPT relies on voltage through being dependent on the socket wattage. EDC is totally exempt from this.

 

[tabascosauz]

Well, if there is any, please provide directions as to how. This normally entails a larger debate as to whether PPT or EDC provides the higher benefit to which I myself base my verdict on, so if you would have just told to undervolt I would thusly state no, PPT is better and even further is EDC. PPT relies on voltage through being dependent on the socket wattage. EDC is totally exempt from this.

Yeah, the surefire way is to set an offset. lol

Is that your test and your graph? I see that it's supposedly made by someone named vpcf90. That's almost the exact same platform as I am running, and I don't seem to get any of those power efficiency gains from reducing EDC or PPT. Since that just eliminated the motherboard variable, that's kinda telling.

I have to run the lower PPT or EDC with stock Vcore or the test would be kinda moot, and I'm not getting any semblance of power savings. Still sounds like a jet engine, and still gets as hot as it does with all stock settings.

 

[Waldorf]

@tabascosauz
please read the first sentence of my post again.

im not saying there is no gains, im saying its NOT because the cpu starts using the highest clocking cores (vs it didnt before) like many thing/believe.
affinity/scheduling for cores can only be "fixed" by MS updates that change windows, not by some added stuff/profile/software etc.

stuff is still heavily tweaked for intel cpus, so i expect this to get better, but with intel still recognizing non intel cpus and treat them like such is another point for massive gains.

 

[mtcn77]

Yeah, the surefire way is to set an offset. lol

Is that your test and your graph? I see that it's supposedly made by someone named vpcf90. That's almost the exact same platform as I am running, and I don't seem to get any of those power efficiency gains from reducing EDC or PPT. Since that just eliminated the motherboard variable, that's kinda telling.

I have to run the lower PPT or EDC with stock Vcore or the test would be kinda moot, and I'm not getting any semblance of power savings. Still sounds like a jet engine, and still gets as hot as it does with all stock settings.

Not, I just research others.
If it struggles to run cool, that is definitely a limited sample issue. You could troubleshoot. However, the general statements wouldn't be jeopardised.
In my experience smartfan is the way to go. It sets a custom bios range where the fan only operates through the prescribed rpm window. 73 degrees Celcius & 88 watts is not hot by my standards and all things considered if there is cpu headroom to spare, having a one stop solution to overclocking through PBO just might be too tempting to resist considering all the hassle of oc-personalisation.

 

[tabascosauz]

If it struggles to run cool, that is definitely a limited sample issue. You could troubleshoot. However, the general statements wouldn't be jeopardised.
In my experience smartfan is the way to go. It sets a custom bios range where the fan only operates through the prescribed rpm window. 73 degrees Celcius & 88 watts is not hot by my standards and all things considered if there is cpu headroom to spare, having a one stop option to overclock through PBO just might be too tempting to resist the hassle of oc-personalisation.

It doesn't really struggle to run cool, it's just that a few degrees happens to be a major acoustic difference on two NF-A9s. I mean, it never really gets up past 74 degrees in anything, now that P95 Small is subject to a handicapped multiplier.

So apparently, in order for EDC 80 to make a difference, all the other PBO settings have to be on Auto. Seems I jumped the gun - that's my mistake. With EDC of 80, it's basically down to the same Vcore and temperature and acoustics range as -0.075V offset.

But there's no power savings and the CB R20 score is still about the same. Going off the graph, I'm not sure if I can recognize a +10pt gain in CB as a real gain; even without changing any settings, that kind of gain is fair game for regular variations.

Also, there's seems to be a LOT more than a 0.006V difference in Vcore between EDC 80 and stock EDC 90 than the graph makes it look like.

 

[mtcn77]

It doesn't really struggle to run cool, it's just that a few degrees happens to be a major acoustic difference on two NF-A9s. I mean, it never really gets up past 74 degrees in anything, now that P95 Small is subject to a handicapped multiplier.

So apparently, in order for EDC 80 to make a difference, all the other PBO settings have to be on Auto. Seems I jumped the gun - that's my mistake. With EDC of 80, it's basically down to the same Vcore and temperature and acoustics range as -0.075V offset.

But there's no power savings and the CB R20 score is still about the same. Going off the graph, I'm not sure if I can recognize a +10pt gain in CB as a real gain; even without changing any settings, that kind of gain is fair game for regular variations.

Also, there's seems to be a LOT more than a 0.006V difference in Vcore between EDC 80 and stock EDC 90 than the graph makes it look like.

Good thing we are on the same page. Now, I spoke to you about Q-fan or smartfan or whatever its name - if you do that, you can do all sorts of things. Be leery about it, if you fine tune your Q-fan for your attention threshold, there is a golden nugget down the line - the cpu does the rest for you. So, there is no exception to not using it. The wee spinner can go up to 5200 rpm! Just set it to 60/30% & 70/60C and I honestly don't hear it over ambient case fans. It idles around 2000-2500rpm, never surpassing 3500rpm - statistically a nonentity. The way I settled through these numbers were I was pushing small fft P95 and just measured how it settled. Notice, on Ryzen you don't even have to undervolt. Just set it and forget it.
PS: there is more than one setting, though. You first do the setting I told you about and then set it on windows fan monitor. The windows settings are 20%/30C, 50%/60C and 90%/70C since it does not allow for a deeper section carved fan lobe.

 

[tabascosauz]

Good thing we are on the same page. Now, I spoke to you about Q-fan or smartfan or whatever its name - if you do that, you can do all sorts of things. Be leery about it, if you fine tune your Q-fan for your attention threshold, there is a golden nugget down the line - the cpu does the rest for you. So, there is no exception to not using it. The wee spinner can go up to 5200 rpm! Just set it to 60/30% & 70/60C and I honestly don't hear it over ambient case fans. It idles around 2000-2500rpm, never surpassing 3500rpm - statistically a nonentity. The way I settled through these numbers were I was pushing small fft P95 and just measured how it settled. Notice, on Ryzen you don't even have to undervolt. Just set it and forget it.

Oh, I always bench with the side panel off and both A9s and A12x25s at full speed. Just to eliminate one more variable. Otherwise, it's small and on my desk, so I try to keep the noise down when I can - the reason why I can't stand maxed out PBO settings, even if there was a performance improvement (there isn't). I intended to use my C14S, which would have tamed the 3700X down right away, but the socket placement prevents that. 2500rpm idle seems a bit high, mine is usually around 1000rpm and goes up to 1500rpm at full load.

I'm not sure if reducing EDC to 80 does much, now. EDC 80, noPBO -0.075V, AutoPBO AutoVcore......they all seem to have the same results in temperature, Vcore, noise, CB R20 results, and power consumption. The settings are sticking, I'm sure of that, but nothing else seems to be different.

The undervolt used to be absolutely necessary. The old F42c BIOS pumped way too much Vcore. Then F50 and 1.0.0.4 probably discreetly changed things, and it appears the undervolt setting doesn't work anymore and Auto just does the same thing. I hate secret behind-the-scenes changes.

 

[mtcn77]

Oh, I always bench with the side panel off and both A9s and A12x25s at full speed. Just to eliminate one more variable. Otherwise, it's small and on my desk, so I try to keep the noise down when I can - the reason why I can't stand maxed out PBO settings, even if there was a performance improvement (there isn't). I intended to use my C14S, which would have tamed the 3700X down right away, but the socket placement prevents that. 2500rpm idle seems a bit high, mine is usually around 1000rpm and goes up to 1500rpm at full load.

I'm not sure if reducing EDC to 80 does much, now. EDC 80, noPBO -0.075V, AutoPBO AutoVcore......they all seem to have the same results in temperature, Vcore, noise, CB R20 results, and power consumption. The settings are sticking, I'm sure of that, but nothing else seems to be different.

The undervolt used to be absolutely necessary. The old F42c BIOS pumped way too much Vcore. Then F50 and 1.0.0.4 probably discreetly changed things, and it appears the undervolt setting doesn't work anymore and Auto just does the same thing. I hate secret behind-the-scenes changes.

One thing you never mentioned is your cpu fan speeds which I find ironic because I had the same issues when using a Zalman CNPS9500 cooler. I hate open heatpipe design starting with it.
PS: looks like you did, my mistake.

If you get where I'm at, you will likely suspect I blame AMD for being too hesitant which would be correct. Where is my single sliding scale at!

 

[tabascosauz]

One thing you never mentioned is your cpu fan speeds which I find ironic because I had the same issues when using a Zalman CNPS9500 cooler. I hate open heatpipe design starting with it.
PS: looks like you did, my mistake.

If you get where I'm at, you will likely suspect I blame AMD for being too hesitant which would be correct. Where is my single sliding scale at!

The U9S is a few degrees cooler than the CNPS9500, but that's about it. There's not much performance left to squeeze out of the 92mm tower; the U9S and D9L have already done that compared to the U9B SE2. I have all three, and it might be time for something larger.

Only one I blame is Gigabyte for their idiotic socket placement, lol. It's always higher than everybody else's, meaning the C14S can't be used.

I'm curious to see what results @lorry might turn out with PPT and EDC reduction. He seems to have a better chip.

 

[mtcn77]

The U9S is a few degrees cooler than the CNPS9500, but that's about it. There's not much performance left to squeeze out of the 92mm tower; the U9S and D9L have already done that compared to the U9B SE2. I have all three, and it might be time for something larger.

Only one I blame is Gigabyte for their idiotic socket placement, lol. It's always higher than everybody else's, meaning the C14S can't be used.

I'm curious to see what results @lorry might turn out with PPT and EDC reduction. He seems to have a better chip.

You could always use air ducting, you still haven't measured how it fares though I know it isn't too efficient, heat pipes don't have particularly good heat capacity vs. conduction.

 

[lorry]

Yeah, the surefire way is to set an offset. lol

Is that your test and your graph? I see that it's supposedly made by someone named vpcf90. That's almost the exact same platform as I am running, and I don't seem to get any of those power efficiency gains from reducing EDC or PPT. Since that just eliminated the motherboard variable, that's kinda telling.

I have to run the lower PPT or EDC with stock Vcore or the test would be kinda moot, and I'm not getting any semblance of power savings. Still sounds like a jet engine, and still gets as hot as it does with all stock settings.

What is this offset then? What does it do and how do you use it / apply it?

The U9S is a few degrees cooler than the CNPS9500, but that's about it. There's not much performance left to squeeze out of the 92mm tower; the U9S and D9L have already done that compared to the U9B SE2. I have all three, and it might be time for something larger.

Only one I blame is Gigabyte for their idiotic socket placement, lol. It's always higher than everybody else's, meaning the C14S can't be used.

I'm curious to see what results @lorry might turn out with PPT and EDC reduction. He seems to have a better chip.

Eh?
"PPT and EDC reduction" ? Thought we were talking about increasing. You have lost me, explain Please?

Just to remind ourselves - there's a 'post your R20 scores' on the AMD forum.
This is a pic from a 3950x user. Look at how far ahead of us his multi core score is, And the compare that amount compared to how little he is ahead of us in single core.
It is multi core where AMD excells, not the single core

 

[tabascosauz]

What is this offset then? What does it do and how do you use it / apply it?
Eh?
"PPT and EDC reduction" ? Thought we were talking about increasing. You have lost me, explain Please?

Just to remind ourselves - there's a 'post your R20 scores' on the AMD forum.
This is a pic from a 3950x user. Look at how far ahead of us his multi core score is, And the compare that amount compared to how little he is ahead of us in single core.
It is multi core where AMD excells, not the single core

The idea of messing with PPT and EDC is that you can achieve similar results to undervolting with supposedly better performance, and allegedly increase your CB score very slightly with a slight decrease in EDC. Your stock EDC is 140, mine is 90. There seemed to be a near-negligible CB improvement with mine set at 80, basically -10. Again, you'll find PPT, TDC and EDC under the PBO menu, and they'll appear if PBO limits are set to Manual. You'll need to have PBO itself on Advanced settings, as before.

However, at least for me, all aspects of behaviour and (otherwise) performance are pretty much identical to disabling PBO altogether and setting a -0.075V offset. When I talk about undervolting, it's always through setting a dynamic-running voltage offset that applies to whatever Vcore figure the CPU is receiving at any given point in time. You'll be able to adjust a negative (undervolt) or positive (overvolt) offset if you set Vcore to Normal as opposed to Auto (just type in N in that field for Normal and A for Auto), in the Core Voltage setting under MIT>Advanced Voltage Settings. But I'm not exactly sure yet if the F50 BIOS has broken the offset functionality, meaning that if you leave it on Auto or Normal, it'll might just do the same Vcore regardless of what offset you set.

Setting a manual Vcore (in the aforementioned Core Voltage field, typing in any [reasonable] number instead of N for Normal and A for Auto will set a fixed Vcore) outside of setting manual fixed clocks is ridiculous, and I can't believe idiot reviewers actually made headlines at launch by setting a 1.0V Vcore. 1.0V is good for all-core manual 3.5-3.6GHz on a decent chip, and that's about it. With the boost still on, Ryzen needs more volts to hit any higher speeds, and if it's not getting those volts, it's not going anywhere past. Simple as that. In the context of undervolting: if you run fixed clocks, set a manual Vcore; if you want boost to work as it is intended, set an offset.

I mean, the 3950X only boosts 100MHz higher. 4.7GHz is more or less the very ceiling of what Zen 2 is capable of on the current process, there's 0 headroom left at that point. Multi-core, they all scale down their multipliers drastically, or they'd all burn up like FX-9590s.

 

[lorry]

The idea of messing with PPT and EDC is that you can achieve similar results to undervolting with supposedly better performance, and allegedly increase your CB score very slightly with a slight decrease in EDC. Your stock EDC is 140, mine is 90. There seemed to be a near-negligible CB improvement with mine set at 80, basically -10. Again, you'll find PPT, TDC and EDC under the PBO menu, and they'll appear if PBO limits are set to Manual. You'll need to have PBO itself on Advanced settings, as before.

However, at least for me, all aspects of behaviour and (otherwise) performance are pretty much identical to disabling PBO altogether and setting a -0.075V offset. When I talk about undervolting, it's always through setting a dynamic-running voltage offset that applies to whatever Vcore figure the CPU is receiving at any given point in time. You'll be able to adjust a negative (undervolt) or positive (overvolt) offset if you set Vcore to Normal as opposed to Auto (just type in N in that field for Normal and A for Auto), in the Core Voltage setting under MIT>Advanced Voltage Settings. But I'm not exactly sure yet if the F50 BIOS has broken the offset functionality, meaning that if you leave it on Auto or Normal, it'll might just do the same Vcore regardless of what offset you set.

Setting a manual Vcore (in the aforementioned Core Voltage field, typing in any [reasonable] number instead of N for Normal and A for Auto will set a fixed Vcore) outside of setting manual fixed clocks is ridiculous, and I can't believe idiot reviewers actually made headlines at launch by setting a 1.0V Vcore. 1.0V is good for all-core manual 3.5-3.6GHz on a decent chip, and that's about it. With the boost still on, Ryzen needs more volts to hit any higher speeds, and if it's not getting those volts, it's not going anywhere past. Simple as that. In the context of undervolting: if you run fixed clocks, set a manual Vcore; if you want boost to work as it is intended, set an offset.

I mean, the 3950X only boosts 100MHz higher. 4.7GHz is more or less the very ceiling of what Zen 2 is capable of on the current process, there's 0 headroom left at that point. Multi-core, they all scale down their multipliers drastically, or they'd all burn up like FX-9590s.

I remember a video by BZ where he went through all the BIOS settings in the gigabyte X470 board and I think he said the max offset the board allowed was 300, I think that was mv but I'm not sure. I have it bookmarked downstairs though.

At some point I'd like to run through an offset with you so that I understand what is required, if that's ok with you?