AMD Ryzen 3000 Builders Thread

[lopri]

PBO seems broken for me. All it does is make my CPU use lower clockspeeds than default in any MT workload regardless of how many threads it spawns. It only worked for SuperPi. Sort of. And default was really close.

PBO does nothing for me, either. Maybe 25~50 MHz higher frequency for a very short period of time but I do not think it lasts more than a fraction of time. Temperature must be very low as well for that to happen.

It surely raises power consumption massively, though.

 

[Kenmitch]

Its 3600 cl15 ram, samsung b-die. I put it in and set the speed to 3600, but aida64 says my memory speed is 51k, and latency, 80 ! Lopri (and others) get 55-56k and 64 or so latency. Thats a big difference. So I use that Ryzen dram calculator, select "calculate safe" and put those timings in. It won't even post. I have to reset CMOS.

Did you check to make sure the voltages are set correctly? Did you look and see if you Asus MB has the vboot dram voltage option which is set on auto by default on my CH6....Auto is a no go as it'll hang on a error code on my CH6.

 

[Elfear]

Its 3600 cl15 ram, samsung b-die. I put it in and set the speed to 3600, but aida64 says my memory speed is 51k, and latency, 80 ! Lopri (and others) get 55-56k and 64 or so latency. Thats a big difference. So I use that Ryzen dram calculator, select "calculate safe" and put those timings in. It won't even post. I have to reset CMOS.

I had the same issue as you with the DRAM calculator. One or two of the sub-timings it spit out wouldn't allow my rig to post at 3600/CL14. I ended up using Snarf Snarf's subtimings along with 1.45V and it posted fine and passed various memory tests.

I'd also look at Kenmitch's suggestion about the boot-up dram voltage.

 

[Aspality]

Gamers Nexus tested the effects of PBO and AutoOC shown in a video here and basically found that with sufficient cooling, you're basically not going to gain anything out of PBO and AutoOC over what Precision Boost 2 already does. And even with insufficient cooling they found that it still didn't do anything.
Link to the article instead if you prefer that.

 

[lopri]

That is pretty much what everyone is finding it. These chips are factory overclocked and there is no headroom for manual anything. A sophisticated algorithm a la Precision Boost squeezes every MHz out of the CPU, completely sucking oxygen out of overclocking fun.

Mark, this setting below ran Prime95 LargeFFT for 12 hours without errors on my setup. My memory sticks are two 16GB sticks.



And new personal record, 64.2. It's all about Infinity Fabric. At 1800 it is not stable, though. There are voltage options in the BIOS that I do not understand and I wonder if any of them can help with Infinity Fabric and/or IMC. Raising vSOC alone did not help.

 

[DrMrLordX]

T These chips are factory overclocked and there is no headroom for manual anything.

That's actually not true for me. While PBO is . . . not doing anything much, I can use manual OC to get more out of the chip. Default/PB2 seems to get me close to 4200 MHz (4180 MHz or so) average with spikes into the 4600 MHz range in games (I did a long test with The Division 2 not so long ago). I can manual OC to 4375 MHz easily and up to 4450 MHz with some effort. So clearly, all-core static OC is the winner for me, if I really want those clocks.

The real question is . . . in games, why do I need the extra clockspeed? I can get a fair number of extra average fps in 1080p medium, but I mostly game in 1440p ultra. Guess how much extra average fps I got in The Divisiion 2's benchmark running with a static OC? 1 fps. Hardly worth the effort. That's with a Radeon VII OCed to 2050 MHz GPU/1200 MHz RAM, too.

Also sheds some light on all those saying, "9900k is still the gaming champion!" Yeah whatever.

 

[lopri]

Well, I probably do not NEED the extra clockspeed but I certainly do want! Considering the boost behavior of my 3700X, I guess the highest all-core OC that is stable (i.e. Linpack/Prime95) is around 4.20 GHz with massive heat output. The thing is, I do not feel like finding it out due to the overall behavior of the chip.

4450 MHz all-core is definitely a golden-chip territory, especially with 3900X.

@Mark: Double check your fclk syncs with the memory clock you pick. At around 1800 (DDR4-3600) you must choose matching fclk frequency manually or it will automatically be halved. You lose good 10 ns from just that.

 

[lopri]

I am getting better at this.

 

[DrMrLordX]

4450 MHz all-core is definitely a golden-chip territory, especially with 3900X.

Depends on what you are trying to do with the chip.

Games? No problem.

CBR20? 4425 MHz is much easier, 4450 takes maybe 1.38v which is potentially dangerous. Still runs cooler than Prime95 though.

Prime95? Ha! Good luck. 4250 MHz is a more realistic goal. Temps are ugly at even 4225 MHz.

The headache is switching clockspeeds to match the workload. Also I think at least some of the 3800x chips are binned higher than mine. And at least you can do bclk OC! I don't know exactly what's making my system take a dump when I do it, but I can't even do 101 MHz bclk without having to shut down and load BIOS safe defaults. It might be my BPX doing it.

The headache is that PBO can't . . . you know, figure things out the way I did and just move clockspeed and voltage around. If I could hand-tailor the boost map, based on current draw and temps, I could probably dial it where I want it. But that's either impossible or simply something I don't yet know how to do with this system.

 

[EXCellR8]

New BIOS for my board (5.80) available so I'll give that a shot tonight. States AGESA 1.0.0.3 ABB but ASRock is suggesting that this is a Matisse-only BIOS

ASRock do NOT recommend updating this BIOS if Pinnacle, Raven or Summit Ridge CPU is being used on your system.

Might try the 3600 memory again at some point but system working well with the Flare X--still getting a weird split-second system stutter at times. Still using onboard sound too.

 

[lopri]

DrMrLordX: Try 102 bclk. 101 doesn't work for me either. Also try VDDP low (0.9 or lower) and VDDG a bit higher at (0.95~1.05) VDDG should be lower than VSOC.

 

[DrMrLordX]

@lopri

I'll give it a shot later. Interesting. bclk holes eh?

 

[DrMrLordX]

@lopri no dice, 102 MHZ bclk was not stable. My machine just doesn't want to "go there". I had the same problem with my 1800x despite choosing a board with an external bclk clockgen chip. I think I would have had to use a SATA SSD connected to the chipset ports (7 or 8) to make that work. My BPX, which was in that old machine and is my backup NVMe drive in this box, is allegedly sensitive to bclk as well.

However, I did discover a few other interesting tidbits.

First off, my board has VRM current settings. Overcurrent to the VRM (increase their switching frequency, basically) along with LLC appears to stabilize the CPU at "bleeding edge" voltage settings.

In CBR20, with Medium LLC (which is LLC2, I think; Low =1, Medium = 2, High = 3, Turbo = 4, Extreme = 5, Ultra Extreme = 6 hahah thanks Gigabyte) I needed 1.31875v vCore to get stability @ 4375 MHz all-core. Actual vcore according to CPU-z was around 1.332.

With Turbo LLC and "Xtreme" VRM current setting, I can use Turbo LLC and 1.29375v vCore to get stability @ 4400 MHz all-core. Actual vcore in CPU-z is a1.32v. VRM temps go up when you crank up VRM overcurrent, which is a problem for anyone that has hot VRMs. During a CBR20 run with "Xtreme" VRM overcurrent, my reported VRM temps around 38-39C. No problem for this board.

Then I did some research into the effects of LLC settings on default and PBO behavior. I ran the system in bone stock config, with only "Xtreme" VRM overcurrent (no reason to avoid it) and different LLC settings to see what would happen in three benchmarks: CBR20, SuperPi 1.5mod XS, and 3DMark CloudGate Physics. I tested Low, Turbo, and Ultra Xtreme.

LLC Low

CBR20 default: 6961 PBO:6961
SuperPi 1M default: 9.796s PBO:10.008s
CloudGate Physics default: 12630 PBO: 12861

LLC Turbo

CBR20 default: 6853 PBO: 7013
SuperPi 1M default: 9.561s PBO: 9.255s
CloudGate Physics default: 13461 PBO: 13213

LLC Ultra Xtreme

CBR20 default: 6792 PBO: 6944
SuperPi 1M default: 9.763s PBO: 9.828s
CloudGate Physics default: 13483 PBO: 12800

Seems like there wasn't a lot of consistency here.

 

[lopri]

I could not get any reason or rhyme out of your LLC results. Maybe that is the point? AMD knows best and these chips aren't meant to be overclocked. I have not even thought of LLC yet, maybe because all-core OC was totally out of my sight. For the most part my time was spent on figuring out memory performance and the voltage settings. I did play with PBO but quickly learned that it was futile. My temporary conclusion is that either

1) These chips have a ridiculous voltage curve where past 4.2 GHz they need an exponential amount of voltage every 100 25 MHz increase. Or,

2) There is another parameter that governs chips clocking behavior, which AMD has not disclosed.

When I was playing with PBO I was almost sure that 1) was the case. Even allowing a bucket of power in BIOS the chip's performance did not budge. It did consume more power, though, which was aggravating to observe. Then I've learned that higher bclk is a possibility, and strangely it brings a higher performance in a somewhat consistent manner. Take a look at this:



All the single core scores below 6100 are produced with bclk 100, and every ST score over 6100 was done by bclk 102. There is also a clear separation in MT scores as well. So whatever the PB does, elevated bclk seems to be able to ignore it at least for the duration of Geekbench 4. (Link) (same is true for GB3)

 

[DrMrLordX]

I could not get any reason or rhyme out of your LLC results. Maybe that is the point?

The only trend I observed is that moving from Normal/Auto -> Low -> Turbo produced generally higher results for PBO. Generally. Defaults got kind of worse, except in SuperPi. I can't explain why Ultra Extreme did so badly. I was "clued in" to the possibility of LLC making a difference when I noticed that LLC Medium vs LLC Normal (off) produced different results. Maybe I should go back and test all the LLC settings for kicks. It's a bit time-consuming though.

 

[EightySix Four]

Anyone that has a 3800x willing to run a little experiment for me? Can you set the TTP (total socket power) to 88 watts and a do a cinebench run? I'm curious how it compares to a 3700x.

 

[Terzo]

I'm finally up and running 1 month later. Picked my computer up from microcenter, the motherboard i had previously exchanged for wasn't working for them either (They also tested a 2200 g in it).

I said screw it, went with the x570 gigabyte mitx and paid for microcenter to install because I was tired of troubleshooting.

Based on my experience, if anyone is looking to upgrade and doesn't have an 2xxx series cpu they can use to flash bios, I recommend saving your self some trouble and going with a 570 motherboard.

 

[PotatoWithEarsOnSide]

I'm finally up and running. Signature updated.
Idle temperature of around 32C and 0.9-0.95v.
Not updated BIOS yet, so still on a pre-release BIOS.
All I've changed thus far is PBO from Auto to Off, and DOCP enabled on the RAM, plus som changes to the fan curves.
Played a battle each of War Thunder 4K Low and 4K Max. Approx 220 vs 115 FPS.

 

[DrMrLordX]

@Terzo

Sorry the B450 board didn't work out for you. At least now you have a working system. I still think B350/B450 seems like a perfect match for budget 3600 systems . . . if they work. x570 is better-supported for now.

@PotatoWithEarsOnSide

What kind of temps are you getting in benchmarks? Prime95 small FFTs? CBR20?

 

[PotatoWithEarsOnSide]

@Terzo

Sorry the B450 board didn't work out for you. At least now you have a working system. I still think B350/B450 seems like a perfect match for budget 3600 systems . . . if they work. x570 is better-supported for now.

@PotatoWithEarsOnSide

What kind of temps are you getting in benchmarks? Prime95 small FFTs? CBR20?

21m of Prime95 Small FFT:
All cores hitting 4.04GHz throughout at a pretty consistent 1.353v, 87-89C

Edit: with BIOS build version 0406 (which is a pre-release BIOS not currently listed on their support page)

 

[DisEnchantment]

Question:
Is anyone noticing a performance degradation with their Ryzen 3K as some of the reddit posts been reporting lately?

I am yet to get my 3900X, hopefully this week, so I am a bit worried. Hopefully those were caused due to some bad settings not silicon issue

 

[PotatoWithEarsOnSide]

Updated BIOS to the latest version, and then had to do a full Windows 10 re-install.
It didn't want to detect my M.2 SSD as a bootable drive.

 

[VirtualLarry]

Updated BIOS to the latest version, and then had to do a full Windows 10 re-install.
It didn't want to detect my M.2 SSD as a bootable drive.

I hate to tell you this now, but... could it have been the default BIOS UEFI/CSM/NVMe settings in BIOS, and maybe you just needed to change them to boot? Like disable CSM altogether, to boot from M.2 NVMe?

Or maybe, it blew away the "Windows Boot Loader" entry in Boot Devices, and the Windows drive didn't re-create it as a failsafe, if you booted the Windows drive directly (CSM mode).

 

[Tsavo]

Question:
Is anyone noticing a performance degradation with their Ryzen 3K as some of the reddit posts been reporting lately?

I am yet to get my 3900X, hopefully this week, so I am a bit worried. Hopefully those were caused due to some bad settings not silicon issue

Nope.

3700x, all stock settings.

 

[DrMrLordX]

21m of Prime95 Small FFT:
All cores hitting 4.04GHz throughout at a pretty consistent 1.353v, 87-89C

mmm toasty. It's interesting that PB2 will let your CPU get that high. Mine won't. Different CPU, but still, the highest clockspeed I can muster with everything bone stock is maybe ~3800 MHz. Maybe slightly higher, around 3812, but nothing significant. I managed higher clocks when I set LLC to Low (as opposed to Auto or Normal), but I kinda wonder if that's clock stretching at work. Temps weren't notably different.

Question:
Is anyone noticing a performance degradation with their Ryzen 3K as some of the reddit posts been reporting lately?

I haven't been reading those posts but I have noticed degradation in only one application: CBR20. Not sure why. I haven't tried to replicate my 7812 run that hard, but the few times I've casually run the application to make sure my clockspeed/voltage were stable, the results haven't been that high. 4425 MHz all-core nets me somewhere in the ballpark of 7700 now. I am still killing my old times in CBR15, 11.5, and 10 though so they're all fine.

Hopefully those were caused due to some bad settings not silicon issue

Probably a microcode issue. People keep updating their UEFI, so things change.