JohnLai I'm interested in your observations on the idle SOC Power usage.

When you tested IF 1800mhz vs higher that consumed more than 10w difference, did you keep the VSOC constant?

My observation that while keeping VSOC at 1.0v, 1800mhz IF vs 1833Mhz only used 1 more watt, but that VSOC is remotely stable, lots of WHEAs lol)

I saved two profiles for myself:
1. mem 3733, IF 1866, vsoc 1.1, idles around 13w
2. mem 3600, IF 1800, vsoc 1.0, idles around 8w

With the first profile, my 5600x allcore boosts less in CB20 than second profile since the 5w of the powerbudget is used by the soc (I keep PBO limits AMD stock values)

Ah yes i forgot u're using 5800x. In the end our findings converge

IMHO idle should be idle i.e. IF should downclock at idle, like how Intel's uncore freq downclocks at idle

AMD really have quite a way to go on fixing these things which might not be apparent to the end user, unless the end user looks for it

JohnLai So I decided to test how low I can go on the vSOC, and it looks like I've settled on 0.95v for now (n hwinfo/zentimings. Bios at 0.96).

Strangely passed 10000% karhu last night with my rams at 16-19-18-39 1.39v, will test tm5 and hci today, and prolly some occt blends to check for WHEAs.

Its only a 1w soc power improvement, but hey I dont have anything else to do lol.

Oh and I dont torrent using this rig nor do I use foobar, so I dont experience the issues that you do...

Edit: ok everything looks good, I'll use this profile as daily driver and observe further for any hiccups



This post has been edited by x800: Mar 1 2021, 05:24 PM

I'm using the gigabyte b550i right now. Yes its really solid, the fullsize backplate, the heavy vrm+m.2 heatsinks. Have a 140mm fan blowing directly from my mesh sidepanel onto the board and VRM temps never went above 52 in a room without aircond (only 5600x tho)

Anyway I'm using an ATX case and have no plans to migrate to ITX form factor. Using this board right now since I got this board 2nd hand for cheap. Anyone with full ATX B550 wanna swap?

This post has been edited by x800: Mar 3 2021, 11:30 AM

IIRC from some readings at OCN, AMD enable uncore oc mode by default (regardless if you turn it on or off) at some point, hence it broke the SOC powerstates making it run at constant freq and voltage and not lowering down during idle. Let me I can find some posts about this and paste it here later.

Nonetheless its true that the effects are more perceivable on vermeer vs matisse, even if that change by AMD was done during matisse era

I think this post #18 was one of it, but the info on "soc powerstates broken because of uncore oc enabled by AMD" is buried in the middle of that post

Yup coz that setting is broken by AMD coz they needed to fix some other things, which was the point i was trying to say

This post has been edited by x800: Mar 27 2021, 07:20 PM

Finally its over.

Anyway JohnLai, here are some results from the testing I did to see the effect of IF clock on SOC power (and the inherent soc uncore mode which is broken even if you set it to disabled):

Testing was done with:
1. Aorus B550i with latest stable F12 (non-beta)
2. Uncore OC mode disabled
2. VSOC/VDDG/VDDP/VRAM/RAM speed/RAM timings set constant
3. Balanced powerplan with latest chipset drivers installed (early Feb 2021)
4. On my Bench OS with very minimal things running in the background

IF clock	Idle SOC Power	CB R20 MC Load Power
1333	~4.4w (screenie)	~5.6w (screenie)
1600	~4.5w (screenie)	~5.8w (screenie)
1633	~8.5w (screenie)	~9.4w (screenie)
1800	~9.0w (screenie)	~9.8w (screenie)

So what I can see are:
1. IF clock above 1600 increases the SOC power by a significant magnitude (this concurs with the reddit post that you linked to earlier, but its actually the IF speed that affects it)
2. IF clock increase between 1300 to 1600, and 1633 to 1800 results in marginal SOC power increase
2. Idle vs Load increase the SOC power by a marginal magnitude (this concurs with the OCN post that I linked to earlier, SOC powerstates are somewhat broken)

Disclaimers:
1. Using a 2 CCD CPU will definitely increase the power figures
2. I did not test by varying the RAM speed, I'm a bit lazy to do that sorry
3. VSOC is keyed in manually instead of Auto, which might be counterintuitive to validate if the Uncore OC mode is really broken
4. I briefly tested IF clocks in between 1333-1600, and also 1633-1800, but the difference was marginal so I didnt record the results or take any screenies
5. Sorry if this tests has been posted here before, but I did it myself to pique my own curiosity and share here

Yup, exactly. The powerstates are broken for Zen3.

For Zen2, well I had some time this afternoon to test out my old 3300x (I forgot how hot this thing is ), cleared cmos, set the voltages and timings, and here are the results.

Again, I used the same test setup (even if this 3300x are not stable using these voltages), but I kept it same as previous results to minimize variables.

IF clock	Uncore OC setting	Idle SOC Power	CB R20 MC Load Power
1333	Off	~6.1w (screenie)	~8.1w (screenie)
1333	On	~6.0w (screenie)	~8.2w (screenie)
1800	Off	~6.6w (screenie)	~9.6w (screenie)
1800	On	~6.5w (screenie)	~9.6w (screenie)

So conclusions based on my testing conditions (other testing conditions may find something else but lets save that for next weekend):
1. Zen3 SOC power increases significantly above 1600IF, whereas Zen2 does NOT have the same behavior.
2. Zen3 SOC powerstates are broken (idle vs load is only 1w difference), whereas Zen2 are working (idle vs load is 2-3w difference)
3. Uncore OC setting in UEFI definitely does nothing (at least using my board bios version)

So, for your case if you really want to reduce the SOC power, either:
1. Go down to RAM at 3200 with IF at 1600 to keep 1:1 (tighten the timings to make up for the lost in frequency of course) OR
2. Find the lowest stable vsoc that you can at the frequency you want to keep it at
3. Cross fingers and hope AMD fixes it

And for your CPU SVI2 TFN not idling...sorry I dont have a 5950x to test...

SVI2TFN, since its kinda confirmed that its powerplan related, why dont you try the custom powerplan posted in the OCN forums here. Try out the BalancedLowPower, see that if does something. I honestly don't see any major change myself, but that's prolly coz I'm using the 5600x and not the other higher core cpus, so i didnt use that powerplan

This post has been edited by x800: Mar 28 2021, 09:45 PM

All in all after all those testings though, soc power consumption doesnt really affect the end user experience. Idle speed temps, same oso not much affect to end user.

But the usb issues that many users are facing does affect the end user experience (bluetooth and LAN issues too for some Gigabyte boards). 1.2.0.1 Patch A beta bioses are out for some boards which supposedly brings the 1.2.0.2 usb fix earlier, but I've read mostly negative feedback.

Now I'm just waiting for desktop cezanne to see what more I can play with in the future..

Itx builds are fun

Buulding customloop inside my friend's nr200 was a whole new experience for me vs my bigass case

Temps was not up to my liking tho...watertemps went above 50 in long gaming sessions even with 2x240rad loop for 5800x+3080...but the owner said its more or less same with other ppl with same config, so up to him la

If you find the CO first, then later put override, it may work but it wont be 100% stable.

The voltage requirements gets higher as frequency increases in a non-linear way. Even if the CO adjusts this v-f curve non-linearly (as it is supposed to be, which is why CO is better than the classic vcore offset), using a CO setting which is stable at 0MHz offset may still not provide enough voltage for a 200Mhz offset.

Yes that WHEA ID 18 event log is useful, but you will still have a risk of data corruption since there may be some read-write operations which is happening when the system crashes with that WHEA error.

I use the corecycler script to check my individual core CO stability. Would take days, if not a couple of weeks to get fully stable, but CO stability testing is less time consuming than fully stability testing RAM OC which is a whole different game IMHO (at least for my case with a 5600x 6 cores only)

This post has been edited by x800: May 17 2021, 05:22 PM