Testing PBO and negative v-core offset with Zen+

[Zerileous]

So with renewed attention to Ryzen and PBO, and an an article claiming that PBO doesn't really mater for Zen2, I set out to discover if it did anything on my system (2600X + x470 Taichi). I know this isn't really relevant to the current discussions, and it's pretty late, but it may still help the Zen 2 builder decide if going from b450 to x470 is worth it (b450 doesn't include PBO for Zen+ CPUs, I'm not sure if it does for Zen2), or decide if going from the 2600 to 2600X is worth it. This will also help those shopping for an X SKU Zen+ and considering buying a motherboard with known problems setting a negative v-core offset

All tests were run on a freshly rebooted system, with background processes killed. Fans were set at 1300rpm, loop deltaT remained between 5C and 10C over ambient. The custom loop contains a total of 720mm radiator (6x120mm). Cinebench R20 was run 5 times for each setting, and the top and bottom numbers were discarded (this was mostly academic, however a couple runs did produce significant outliers, but rather than calculate this, I just did the same for every run). The three remaining values were averaged.

Starting at "factory optimized defaults" I got 3075.

Enabling XMP to 3200MHz showed one of the biggest performance improvements to 3133 (shouldn't be a surprise).

Next I tested Vcore offsets from +0.05V down to -0.075V. At -0.075V the score improved to 3146. At -0.1V the system crashed loading Windows. At +0.05V the score plummeted to 3103.

Returning voltage to Auto, Determinism was set to Performance and PBO was enabled. This gave a minor improvement over the XMP Auto Voltage score of 3147. At -0.05V more improvement is seen to 3169, while at -0.075V a result of 3181 was achieved. This latter value proved to be stable running Prime95 Small FFT for 3 hrs, however was unstable during moderate and single core load spikes (i.e. resuming from idle, or running Cinebench R20 single core mode), so the offset was decreased (or increased) to -0.0625V, which has been running stable. I did not go back and get a score at this setting.

Conclusions: Memory speed makes the biggest difference. Negative vcore offset and PBO each make a small difference in performance. Combined they can make as approximately much difference (in Cinebench R20 scores) as setting memory from stock to 3200MHz. Of course YMMV given cooling etc. compared to a heavily raded custom loop.

 

[ShrimpBrime]

Questions:
1. Was the VRMs stabilized to help prevent v-core fluctuations?
2. Was winders optimized via performance or power saving modes?
3. Which operating system was used for testing? (this may have core usage impact)
4. Have you tried a positive offset for any testing yet? (curious if additional voltage helps increase scores and limit leakage)

Good testing though, the results seem accurate.

At +0.05V the score plummeted to 3103.

Ah, I see scratch #4 I suppose.... At what temp did it seem to start throttling? (stock & OC?)

Additional Comment:
I noticed with a too large negative offset (too low of a v-core), the throttle happens at the reference clock (FSB) and multi stay put.

 

[Zerileous]

1. I'm not sure what you're asking. No modifications were done to the motherboard. Given the VRM on the Taichi could handle 16 cores, I don't believe that VRM stability was a factor. Vcore was set to Auto for all testing however, so it did fluctuate according to whatever the soup of precision boost overdrive, XFR2 and Precision boost 2 determined.

2. I was really puzzled by finding my "winders" for a bit. Ryzen Balanced was the windows mode.

3. Windows 10, will update with the version 1903, good point though.

4. I don't remember the specific temp, but I believe it was under 85C the whole time, I definitely didn't see over 90C. I could re-run it and check. My original plan was to SS my HWiNFO64 screen after each set of runs, but then I got in a rush and skipped it.

I never noticed my bclk/fsb dropping, however all of my problems with too low of an offset resulted in a freeze/crash. At -0.075V saw the best all core performance but single threaded CB runs would cause a crash.

 

[ShrimpBrime]

I suppose my reference to bus clocks falling is generally from static overclocking, so the XFR/PBO and CPB are all disabled. Mainly I was testing the low voltage point, curious to see if the typical low volt hang would occur as it did in previous platforms often times called a hang up.
So I was experiencing the reference clock dropping as low as 89mhz, it was quite impressive to say the least. And all this with temps in check say 70c and the load didn't need to be heavy.
Tested with performance mode on and VRMs set to extreme. That way I could keep a stable voltage and monitor for either leakage or stability issues from too much negative offset. Turns out the Cpu is smart enough to compensate not enough voltage. Consequently, I did have lower cpu temps.

But now I'm using a different board. The above was with a B450M-A and now have installed ROG B450-I Gaming. They are two totally different creatures. Firstly, my testing will take longer over all. There's quite a bit more bios candies on the ROG. The new board has also allowed me to run 1500+ effective memory clocks while the B450M-A was very unstable past 1333, however I was able to run Cas 14-13-13 and seemed to perform pretty decently with the Cosair memory. But that has changed and now get a much higher effective clock at 16-17-17 and has helped performance between 1-3% depending on the benchmark.

Perhaps I will try to demonstrate the reference clock dropping. I'll try this at 3.7ghz @ P-state voltage which is 1.212v. I'll use a negative offset and see how this ROG board reacts. (So far only testing for stability with Mem clocks and timings the past week.) I'd imagine to have a similar outcome as the B450M-A.
Am curious how the CPB.PBO.XFR act once I get down this road you're currently testing on. I hope for nice results like yours.

 

[ShrimpBrime]

Ah here you go. The ROG board seems to handle the static p-state low v-core FSB droop a tad better. VRMs are on auto this time here.
Negative offset at 0.02250v / performance mode / Boost and c-states and SMT all disabled. It droop to average of 97 Mhz.
SS taken at load temp.



And this screen shot was all stock negative offset of 0.02250 and experienced a droop to 99.77Mhz which brought the all core boost clocks from 4Ghz to 3916 and when temps hit 71c the fan went full on. At the end of the stress test of full load, a single core boost After idle temps the cpu will hit 4350mhz testing using PiMod 32m. (good for memory testing too). Single core boost Is 1.412v or if enabled VRM stage 4 of 5 will do 1.450v automatically.

Nobody else gets these FSB droops? Drives me nuts trying to figure it out lol.

 

[Zerileous]

Interesting, I think I've seen that before but not sure if it was on SS or my own system. I always assumed it was just variability in the clock gen from system to system but I can take a closer look. I don't think my board has any settings for the VRM other than LLC, which I left to auto.

 

[trents]

If I'm understanding correctly what I've been reading here and other places, the TDP designations and boost speed claims from AMD for the newer Ryzen CPUs is subject to some significant caveats. In reality, boost levels seem highly dependent on cooling and quality of motherboard components, even at default settings.

 

[ShrimpBrime]

Not sure Trents. I seem to have similar experince on 2 different boards however both mid ranged of the B450 class and almost seems to vary mostly on temp.

VRM count so far has done no significant improvement for how the cpu acts.
Voltage seems to really matter when it comes to throttling. Same with temps.
My experince with throttle seems much different from most other people. Being at 70c I find this a thermal point my particular 2700x does not like to pass. It will throttle a little going from 4ghz all core down to 3.8ghz with some change.

Cooling is such a large variable between systems, its hard to really tell if that is a big factor.
I think some people mistake an all core boost with a single core boost and expect PBO to clock as high as the XFR which is meant for single core clocks. The FXR is extending past PBO where mine will do happily 4350mhz and always below 70c during this action. While a lot of the other threads are idle at the 2.2ghz p-state while c_states are enabled.

Also, even with say... Sub ambient temps, we are not seeing any increase in end OC for anyone.
I can do at the very best 4.4ghz on 4 cores and it takes all of every bit of 1.515v to do so. Now this is higher than stock XFR PBO voltage range for my 2700x which doesnt seem to pass 1.412v for any reason on automatic, I assume that is the highest boost voltage available, however is Not listed as a P-state where past gen AMD cpus included the boost clocks as a P-state. Well obviously this is no longer true.
I did not see an increase in voltage while automatically boosting cores even at the -20c range. Since I havent used LN2, I cannot comment that -170c would make auto volting go any higher.

Since Boosts are not listed as Pstates I do not consider them any type of base clock past the 3.7ghz pstate as the boost clocks will greatly vary depending on temo and voltage used.
Therefor there seems to only be a max all core boost while anything from 3.7 to 4ghz is just a boost range.

Now I have used the term AI while the cpu has sensors and algorithm designed for controlling itself in certain environments. Obviously at 95c you are not going to get a max boost from your chip while I experience a FSB droop on low voltage manually applied.
I for one can run 4.1ghz which is 100mhz past the automatic all core boost with volts on auto equals 1.412v and is very stable and temps are nearly identical. This may be just sheer luck on my part.

 

[trents]

How Ryzen handles power and manipulates clocks seems mysterious.

 

[ShrimpBrime]

Yes and no. I understand that AMD uses SB-TSI (Sideband Internal Temperature Sensor Interface.) and has been doing so for well over a decade. This in communication on 7-bit SMBus must follow TempHighAlert and TempLowAlert. With processor models 00h-2Fh the TempHighAlert is 70c. This does not mean it will throttle the processor, but has the potential to throttle under a stock configuration meaning no changes where done in the bios.

So an OEM computer with say a 2600X and temps exceeding 70c will not only Alert, will also follow parameters to either stay or lower this temperature. This could also be in part of OEM bios parameters where that threshold could be lowered to only 60c. At 70c the Fan should be at 100% operating state unless otherwise changed by user. SB-TSI CPU temperature readings and limit registers encode the temperature in increments of 0.125 from 0 to 255.875. The high byte represents the integer portion of the temperature from 0 to 255. One increment in the high byte is equivalent to a step of one. The upper three bits of the low byte represent the decimal portion of the temperature. One increment of
these bits is equivalent to a step of 0.125.

I figure the trigger inputs and low level output voltage (at 3 mA sink current) are alerted during my manual under-volting and the Cpu simply compensates for it, with this the down-clocking of the reference clock.

As it gets more interesting because it happens at any temperature even if I was running only 3.5ghz and sub 60c load temps. (core and thread count need not apply) I guess end result that too much under-volt ruins performance (but gains efficiency if where able to maintain clock) during manual overclocking.

My problem is that I don't see users having this condition during their overclocks because they strictly use PBO and tweak that instead.
So my curiosity is further sparked by this thread.

My Apologies.

 

[Hotrod2go]

I fiddled with this concept the OP is stating on the rig in my sig & it made no effect whatsover on PBO performance other then from leaving Vcore controls on auto in bios.
Maxed out all the options in bios for manually configuring PBO boost potential & no diff in PBO performance than from leaving them on "Enabled".

The PBO algorithm has been set to cover a broad range of diff quality silicon so as to prove to customers that the claimed PBO clock speeds are going to happen & not disappoint consumers.
This is obviously AMD's way of ensuring that false advertising is not going to happen & tarnish their image.

This is why I turn it off now completely in my bios & just go with a manual OC. My silicon will have slightly diff characteristics when running PBO than to another user of 2600X chips with PBO turned on.
I was regularly getting 4.25Ghz boost on ALL cores during gaming sessions as measured by HWinfo. This was a maximum level, no matter what bios or chipset drivers I used since setting up my rig.

 

[Zerileous]

Windows version 1903. I will leave HWiNFO64 running for a while and see if I ever see a minimum bus clock show up. Right now its 100.0MHz.

I fiddled with this concept the OP is stating on the rig in my sig & it made no effect whatsover on PBO performance other then from leaving Vcore controls on auto in bios.
Maxed out all the options in bios for manually configuring PBO boost potential & no diff in PBO performance than from leaving them on "Enabled".

The PBO algorithm has been set to cover a broad range of diff quality silicon so as to prove to customers that the claimed PBO clock speeds are going to happen & not disappoint consumers.
This is obviously AMD's way of ensuring that false advertising is not going to happen & tarnish their image.

This is why I turn it off now completely in my bios & just go with a manual OC. My silicon will have slightly diff characteristics when running PBO than to another user of 2600X chips with PBO turned on.
I was regularly getting 4.25Ghz boost on ALL cores during gaming sessions as measured by HWinfo. This was a maximum level, no matter what bios or chipset drivers I used since setting up my rig.

Negative vcore offset has been broken so I'm not surprised that it made no impact on your performance. I am curious though how you measured this.

A manual OC is great Hotrod2go. I chose this method because many games don't use all the cores, and in those situations a single core can boost higher than I can manually OC all of the cores. I also cannot manually OC all of the cores much higher than PBO with voltages that I am comfortable with.

What games were you playing and how much were they loading the cores? Cinebench loads every core 100% with a rendering task. I'm sitting at 4.25Ghz all core right now, but the total CPU usage is only 3%. I'm actually not sure what's keeping it from down clocking since they aren't doing much, but my point is that at light loads this is not unusual. 4.25GHz all core during a all core benchmark or stress test is a different story.

All told, I never claimed to have a better than average sample, so if you're doing much better then congrats. Just curious about the details is all.

 

[Hotrod2go]

Windows version 1903. I will leave HWiNFO64 running for a while and see if I ever see a minimum bus clock show up. Right now its 100.0MHz.



Negative vcore offset has been broken so I'm not surprised that it made no impact on your performance. I am curious though how you measured this.

A manual OC is great Hotrod2go. I chose this method because many games don't use all the cores, and in those situations a single core can boost higher than I can manually OC all of the cores. I also cannot manually OC all of the cores much higher than PBO with voltages that I am comfortable with.

What games were you playing and how much were they loading the cores? Cinebench loads every core 100% with a rendering task. I'm sitting at 4.25Ghz all core right now, but the total CPU usage is only 3%. I'm actually not sure what's keeping it from down clocking since they aren't doing much, but my point is that at light loads this is not unusual. 4.25GHz all core during a all core benchmark or stress test is a different story.

All told, I never claimed to have a better than average sample, so if you're doing much better then congrats. Just curious about the details is all.

Using same version of win 10 as you & HWinfo v6.10.
How do you know negative vcore offset is broken on my board? there was no vcore offset from MSI with this board until recent bios updates.
I only play Betheada games so the creation engine is regularly run on my system when gaming. Doesn't matter if I just launch HWinfo, does it's thing & loads onto the desktop, it reports 4.25GHz max clock speed immediately even with just that app open & nothing else. Same clock speeds of 4.25GHz when either benching or gaming, doesn't matter. It's like HWinfo has already calculated what the PBO limit is with my 2600X. It's done this consistently since I setup this mobo+cpu combo last year, updating bios, windows, HWinfo or chipset drivers makes NO diff to the PBO boost. Nor does manually upping the controls for PBO in bios.

 

[ShrimpBrime]

Do tell how you fellas manage a PBO ALL core boost to 4.25Ghz please??!!!

I seem to have PBO settings on this ROG board that where not available on the other board.
Something I haven't tried playing with yet as this is still pretty newly set up.
This 2700X with PBO settings on auto is only 4ghz.

 

[Hotrod2go]

Do tell how you fellas manage a PBO ALL core boost to 4.25Ghz please??!!!

I seem to have PBO settings on this ROG board that where not available on the other board.
Something I haven't tried playing with yet as this is still pretty newly set up.
This 2700X with PBO settings on auto is only 4ghz.

Only going by what the ever popular HWinfo tells me. I tried with HWmonitor during gaming & got screenshots to show It got 4.6GHz boost on all cores!
What app is 'honest' here? you tell me...

 

[ShrimpBrime]

Well.... I have HWinfo but seldom use it.

Cpu-z and Gpu-z are the 2 main programs I use. Quick easy simple and often used.
(I use GPU-Z for Cpu temps as well as monitoring my video card at the same time)

So my settings = .... which I've not taken off auto yet besides disable.

PBO Auto/Manual/Disabled

PPT - Limit (auto)
TDC - Limit (auto)
EDC - Limit (auto)
PBO scaler (auto) - Manual 2x through 10x.

Typically manual OC here. Didn't have these settings on the previous board it was just auto or disabled.

Thanks for the help.

 

[Hotrod2go]

Well.... I have HWinfo but seldom use it.

Cpu-z and Gpu-z are the 2 main programs I use. Quick easy simple and often used.
(I use GPU-Z for Cpu temps as well as monitoring my video card at the same time)

So my settings = .... which I've not taken off auto yet besides disable.

PBO Auto/Manual/Disabled

PPT - Limit (auto)
TDC - Limit (auto)
EDC - Limit (auto)
PBO scaler (auto) - Manual 2x through 10x.

Typically manual OC here. Didn't have these settings on the previous board it was just auto or disabled.

Thanks for the help.

I only use HWinfo as it seems to be the "official" tool to use these days for measuring hardware performance in real time.
NOT saying GPU-Z or CPU-Z is useless but as an all in one app I can see why it's popular.

I have PPT, TDC & EDC maxed out to 999 (IIRC because I'm not on that machine atm) these some others I forget the names of now that max out to 255 but have pushed all those as far as they can go & it made no dff to my PBO clock speeds when system under load.
Use to have PBO scaler option in earlier bios but MSI dropped it on my board for some reason with later bios revisions.

 

[ShrimpBrime]

Wait WUT?!

So they (we) use an "official" tool that gives false readings and then someone happens to take a timely screen shot and that's now accepted/acceptable?? wow.....! Your All core boost is higher than the single core XFR for that 2600X. I believe this to be impressive if SMT is enabled.

In order for me to pass IBT and OCCT / Prime (2 hours) at 4300mhz, I have to disable SMT and increase v-core a considerable amount. The chip is totally unstable past 4.2ghz with SMT enabled.
CPB/PBO = 4ghz. Never have gotten the chip to exceed this speed with any setting other than manual all.

 

[Hotrod2go]

@ShrimpBrime, you gotta remember that being p95 or OCCT stable is different from being stable for gaming purposes.
I've never heard or seen any PC game that maxes out every core\thread 100% consistently, unless it is severely poorly coded.
I can bench 3DMark (Firestrike)& PCMark10 with 4.4GHz all cores & SMT enabled.

Here's a screenie from about a yr ago with earlier bios & chipset drivers of course, but with my older kit of ram @ 4.3GHz 1hr blend

& here's IBT @ 4.35GHz 10 runs on standard but with 8GB system ram

 

[ShrimpBrime]

Holy shi...!! Now that's nice. 1.35v even. Low 60f idle temp with a load temp of only 63c!!

Now that's max efficiency my friend!!