• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

Ryzen Owners Zen Garden

Remember that we don't have the CPCC2 thingy that chooses the fastest cores yet, a small performance loss like 770 dropping to 730 is reasonable - dropping to 187 however, was f*cked

There is a missing feature from the CPU's that needs a chipset driver. That small performance loss was well known and expected (the large one was not)

It feels like this patch partially/fully fixed the preferred cores already. I know it wasn't in the patch notes, but AIDA (the worst offender pre-patch, used to literally reverse the CPPC hierarchy) is now back to Win 10 behaviour when loading the cores.

You know what's funny? I should just bench AIDA with half a billion things open, instead of being a tryhard and doing the usual pre-bench cleanup. It has a habit of benching better with clutter, this is like the 100th time I've seen this phenomenon. Maybe AIDA just hates the CPU being idle:

win 11 fixed l3.png


L3 latency still slightly higher than Win 10, but it looks like the patch does work to some extent. Between AIDA being AIDA and Win 11 being Win 11, some variation is probably normal.
 
It feels like this patch partially/fully fixed the preferred cores already. I know it wasn't in the patch notes, but AIDA (the worst offender pre-patch, used to literally reverse the CPPC hierarchy) is now back to Win 10 behaviour when loading the cores.

You know what's funny? I should just bench AIDA with half a billion things open, instead of being a tryhard and doing the usual pre-bench cleanup. It has a habit of benching better with clutter, this is like the 100th time I've seen this phenomenon. Maybe AIDA just hates the CPU being idle:

View attachment 221152

L3 latency still slightly higher than Win 10, but it looks like the patch does work to some extent. Between AIDA being AIDA and Win 11 being Win 11, some variation is probably normal.
yup - your numbers smash mine, but you also have way faster RAM

W11 is now within the ballpark of where it should be, and when we know more patches are coming via the AMD chipset driver no one should be concerned
 
I don't have a retail copy of AIDA so... Trial version it is...

I am impressed by how far Ryzen has come compared to where it was. My system is a prime example of that.
 

Attachments

  • cachemem2.png
    cachemem2.png
    239.6 KB · Views: 88
I will gladly take a [relatively] prompt fix, but I'm not so sure whether it's actually fixed, still looking kinda sus.........

Here's your shots; Win10 vs Win11 fixed:

3800cl14-2t-54-6-png.221088
3800cl14-kinda-fixed-win-11-png.221087


Looks like the L3 write performance is still lacking; only about 60% of where it should be.
 
Last edited:
You are on the beta channel, yes?
(Your ram is faster than mine, so ofc my results are slower)

1634549052697.png
 
Last edited:
Wait, viper posted i left my email in somewhere... the heck did i do that

Thanks for fixing if you did, bro
 
Wait, viper posted i left my email in somewhere... the heck did i do that

Thanks for fixing if you did, bro
It was in the attachment file. I reported it.
 
Haven't checked yet, but i think this thing not much smaller than my HAF X. Pic:

20211018_165147[1].jpg


Will have system up & running by end of December, unfortunately: water cooling system (EKWB) is one of the reasons. :)

P.S. I still love my HAF X, it's just i love this case even more: it's so modular it's a bit ridiculous, actually. :)
 
Bah..

I was excited to grab the new version of Aida64.. my key is still good but I have to buy a new key to use the new version. Apparently 10 years is long enough :kookoo:

I had an Engineer key as well, still good but not on the new version.

So I was running through the tests because seeing Trial Version all over the place made me feel like a noob :wtf:

So I got a bit moist when I got to the AES test..

Capture.PNG


But then afterwards I checked my temp and holy carp it hit 90!! It has never done that before with pbo.. the 90 wasn't for AES, it was for FP32 and 64 Ray trace :oops:

Capture2.JPG
 
The secret sauce for intel is to just lower the clock speed under those workloads :P

dont forget you can tweak PBO, including setting a max temp limit
 
dont forget you can tweak PBO, including setting a max temp limit
I did actually forget about setting the temp limit, normally its not something I need to worry about :D

Maybe I need to put my old fans back in..

As I look into my case and realize my GPU is too long. Hmm.. it might work..
 
I mean yeah, it's a silly little thing but sometimes you just say boost away little chiplet... but you got a speed limit of 80 obamas per football field there, little friend
(translated from celcius)
 
We use both Imperial and Metric here, not even a tasteful blend.. if we get a blueprint that's in metric, we convert it because we measure in inches. Temps are Celsius, I buy 4 liters of milk every week, its 3 kilometers to work from here. But I watch American TV, listen to their music, eat their food, make stuff for them, buy stuff from them..

But! a bit of a detour towards your neck of the woods..

I did make Roo bars for a couple of years when I was in my early 20s.. those guys were from New Zealand. I learned to make em your way.. waxing that blade is scary as fark. It will either put hair on your chest, or give you stiches or worse. When you see that part of your glove between your thumb and index finger turn into a puff.. poof haha. I drank a lot back then. New, more genteel fellas would come and complain the work was for convicts before they left lol. Good times. Lots of overtime..
 
I mean yeah, it's a silly little thing but sometimes you just say boost away little chiplet... but you got a speed limit of 80 obamas per football field there, little friend
(translated from celcius)
I think even 85 obamas might not be that bad. We need more Obamas after 4 years of Trump...
 
Soc to 1.15, DRAM to 1.4v

How is this remotely stable with 64GB?
(PBO is off while i tested this stupid idea, hence lower L3 cache numbers, etc)
1634727496366.png
 
Had to go back to Win 10 just for a sanity check to make sure I wasn't tripping. The L3 cache fix doesn't work, writes are still all over the place. The new chipset drivers with the preferred cores doesn't change anything.

The screenshot at the top of p.165 is the only run I've ever managed on Win 11 where writes aren't still down in the gutter around ~200-400GB/s. Yes, AIDA is unpredictable but not when it comes to L3 perf on this CPU, every run on Win 10 is as it should be.

Just to be clear,

However, it also doesn't look like it's only Win 11's fault. I'd put my thumb on either AGESA or Asus BIOS writers being part of the problem - simply downgrading back to clean Win 10 still kept my MT performance somewhat low (although ST came back as soon as I left Win 11). CPU-Z ST on Win 11 simply hit a wall at 675 and refused to go further despite hitting the same effective clocks.

Just to rule out the board, I went to the 3601 BIOS (AGESA 1201A) and now am able to match my former Unify-X (AGESA 1201B) with my 130/90/120 profile. It is a little more temperamental than the Unify-X, but the performance is there. Amusingly, Win 11 performance was so low that I scored higher on MT @ 130W with 10, than I did on MT @ 142W on 11.

Win 11 vs. Win 10:

3800cl14 kinda fixed win 11.png
win 10 21h1 after 11 54.6.png


Unify-X on 10 @ 130W vs. Impact on 10 @ 130W - proof that the board is clearly performing as it should:

5900x reduced 130w highest st 687.png
5900x 130 90 120 impact.png
 
So if you smash away at the L3 cache write specifically (double click retests) in W10 its consistent, but W11 its erratic?
 
So if you smash away at the L3 cache write specifically (double click retests) in W10 its consistent, but W11 its erratic?

Correct. Weird, isn't it? If this was a question of Win 11 loading the wrong cores, then the CPPC2 fix wasn't even needed because Win 10 loads the same cores as pre-patch Win 11 during the L3 portion. During mem/L1 I observed it loading some non-ideal CCD1 and CCD2 cores, but during L3 it always pushes Core 0 hard ~5.0GHz - same on W10 and W11.

Just ran the L3 portion 20 times in a row on W10, and it's always 770/700/760/10.5 without missing a beat (sometimes writes dip a bit to 670 and other times up to 750, others up to 800).

I don't like to entertain conspiracy theories but this is getting ridiculous. MS has had Zen 3 performance down PAT for months. Why now?

Unless it was AMD sabotaging their own AGESA again. Which seems less likely but is not impossible, Reddit collectively seems to hate AGESA 1.2.0.4 so far.

Edit: leaning towards an even split of MS'/AMD's/Asus' fault. On BIOS 3601 and W10, CPU-Z results are now as consistent as on Unify-X, no more fluctuations. Also the CPU-bound game that was mysteriously clocking low on W11 is also back to its usual self, it wasn't the game after all. No more of that offload-half-the-crap-onto-CCD2 BS.
 
Last edited:
I wanna test with like process lasso, or disabling SMT entirely and see if that alters things
I still have a suspicion its core bouncing to the SMT threads

Disabled SMT, performance went to *shit*
1634860636143.png


Rebooted, SMT on and virtualisation turned on (same menu)
1634861248705.png


Exploit protection exemptions set, no change. I still get high results, but repeated smashing of the test makes the numbers go wild.
Giving it a pause before doing again, fast.
 
Last edited:
I wanna test with like process lasso, or disabling SMT entirely and see if that alters things
I still have a suspicion its core bouncing to the SMT threads

Disabled SMT, performance went to *shit*

Rebooted, SMT on and virtualisation turned on (same menu)

Exploit protection exemptions set, no change. I still get high results, but repeated smashing of the test makes the numbers go wild.
Giving it a pause before doing again, fast.

Yeah I run with default protections, but no performance-hogging virtualization or hardware-based security (I have both trusted computing and fTPM turned off for W10, and I think either Secure Boot or further setup of fTPM is required for the Hardware Security page).

There is some natural variation in the results on W10 (~100GB/s) but nowhere near what W11 does, and I never need to give it a short break before testing again. Just testing back to back and it never misses a beat. I can give it all the break I think it needs in W11 and the numbers will be equally on crack for 49 out of 50 runs.

SMT doesn't work properly on my 5900X, even on W10. It breaks some readings in HWInfo and performance is in the shitter in MW2019 (it'll change your whole perception of "stutter").

I wonder if I can dual boot 10 and 11 on the same drive? Can you expand a GPT partition with free space after you've shrunk it to make room for another partition?
 
I've just tried the fix for 11.. gotta say I am fairly disappointed. I am creating a 10 install right now. Bummer do0d
 
So if you smash away at the L3 cache write specifically (double click retests) in W10 its consistent, but W11 its erratic?

Okay, retested with 3601 (1203A) this time, with L3 fix and CPPC fix. Performance is better and more predictable. Unfortunately, the invisible ST wall is still there. Inexplicable since it hits the exact same ~4.918GHz clocks in both W10 and W11.

I've taken the time to make the W11 install so clean and barebones it hurts - it can bench, search google and that's pretty much it.

Less unpredictable L3 bandwidth results (still low sometimes but 9 out of 10 are kinda okay), but now the inconsistencies manifest elsewhere. 15ns is not normal, constantly fluctuates to 11-15ns:

win 11 3601 unpredictable l3.png


10 vs patched 11:

5900x 130 90 120 impact.png
win 11 not fixed st.png
 
I've just tried the fix for 11.. gotta say I am fairly disappointed. I am creating a 10 install right now. Bummer do0d
It drove me to Linux. Ironically despite wine translation, AIDA64 registers great latencies there.
 
Gave up messing with Dram Calculator. Used XMP timings (the one listed in the profile name, where tRC is 58 instead of 84 that's set once you enable it) with everything else set on auto. DRAM Voltage 1.4V, SOC 1.1V, VDDG 1.05V, VDDP 1.0V with OC on my Ballistix
1635340283038.png

It didn't crash during 2.5h of MemTest64 (previous attempts at 3733 with Sport LT's 3200 lead to reboot within first 20 minutes) so it looks like it's stable. Long gaming session during the weekend will show if it's actually stable.
 
Bah..

I was excited to grab the new version of Aida64.. my key is still good but I have to buy a new key to use the new version. Apparently 10 years is long enough :kookoo:

I had an Engineer key as well, still good but not on the new version.

So I was running through the tests because seeing Trial Version all over the place made me feel like a noob :wtf:

So I got a bit moist when I got to the AES test..

View attachment 221531

But then afterwards I checked my temp and holy carp it hit 90!! It has never done that before with pbo.. the 90 wasn't for AES, it was for FP32 and 64 Ray trace :oops:

View attachment 221534
190W is a bit much. I would lower the limit a bit.
 
Back
Top