Ryzen Owners Zen Garden

[tabascosauz]

The latency is good but the load on the imc is probably too much hence the really poor L3 perf.

I've spent some time thinking about it and it's gotta be something to do with the IF. But I honestly don't know why it's behaving this way because I know if I give it 1.2V it'll do 2200MHz stable IF.

IMC is also a possibility, but I thought about it and 4200 isn't even close to maxing out the Renoir UMC. Just to make sure I'm not insane, I tried to go back through some of BZ's videos with his 4750G to make sure he isn't getting the same symptoms when he's taking his Rev.B to 4500-5500MT/s - unfortunately he likes Geekbench and really never runs even AIDA for the memory test, let alone the complete cache test.

On the topic of Geekbench 3 I did score a hair under 8800 memory score today which is pretty respectable. I'm also getting the gains I should from memory speed in the games I play on this comp, so either there is something I'm missing here or AIDA is just unpredictable (I know it is, but 300GB/s seems a bit much).



It's not BankGroupSwap either - the benchmark penalty for having it off still holds true and I don't notice any gains in game with it off. AIDA is still just as unpredictable with BGS, BGS APU or BGS ALT.

It doesn't appear to be VSOC either.

 

[Mussels]

Reebooted, set to 3800/1900 - no WHEA thanks to latest AGESA

DRAM calc does not support this RAM, its pretty useless for anything not B-die

 

[ViperXTR]

Uuh, my L3 cache score seems pretty low too :O

 

[thesmokingman]

I've spent some time thinking about it and it's gotta be something to do with the IF. But I honestly don't know why it's behaving this way because I know if I give it 1.2V it'll do 2200MHz stable IF.

IMC is also a possibility, but I thought about it and 4200 isn't even close to maxing out the Renoir UMC. Just to make sure I'm not insane, I tried to go back through some of BZ's videos with his 4750G to make sure he isn't getting the same symptoms when he's taking his Rev.B to 4500-5500MT/s - unfortunately he likes Geekbench and really never runs even AIDA for the memory test, let alone the complete cache test.

On the topic of Geekbench 3 I did score a hair under 8800 memory score today which is pretty respectable. I'm also getting the gains I should from memory speed in the games I play on this comp, so either there is something I'm missing here or AIDA is just unpredictable (I know it is, but 300GB/s seems a bit much).

View attachment 185764

It's not BankGroupSwap either - the benchmark penalty for having it off still holds true and I don't notice any gains in game with it off. AIDA is still just as unpredictable with BGS, BGS APU or BGS ALT.

It doesn't appear to be VSOC either.

If you do a test from say 3600-3800-4000mhz, the cache perf seems to go down as you go up in speed. At least it does on my rig and I assume others but don't think this has been a big discussion point so it's vague. Try it and see.

View attachment 185766

Uuh, my L3 cache score seems pretty low too :O

Btw you wanna be on the latest aida version for Ryzen 5k support.

 

[biffzinker]

Btw you wanna be on the latest aida version for Ryzen 5k support.

His screenshot at the bottom shows 6.32.5600.

Version: 6.32.5600 stable (Dec 15, 2020)

Release notes:

• AVX2 and FMA accelerated 64-bit benchmarks for AMD Zen 3 based Ryzen 5000 Series processors

 

[thesmokingman]

His screenshot at the bottom shows 6.32.5600.

Version: 6.32.5600 stable (Dec 15, 2020)

Release notes:

• AVX2 and FMA accelerated 64-bit benchmarks for AMD Zen 3 based Ryzen 5000 Series processors

Doh, my bad. I must have read the numbers wrong.

 

[tabascosauz]

If you do a test from say 3600-3800-4000mhz, the cache perf seems to go down as you go up in speed. At least it does on my rig and I assume others but don't think this has been a big discussion point so it's vague. Try it and see.


Btw you wanna be on the latest aida version for Ryzen 5k support.

Down to 3600, it's still doing it. I think imma just write AIDA off as a lost cause and go back to my usual profiles (which honestly it might as well be lol, hopelessly synthetic).

Interesting observation on the cache. I haven't really noticed a general decay in L3 performance up to 4600, just the weird 1/2 L3 read half the time. The L3 R/W/C numbers are so variable all the time I don't pay much attention to them until they pull a stunt like this.

 

[biffzinker]

Down to 3600, it's still doing it. I think imma just write AIDA off as a lost cause and go back to my usual profiles (which honestly it might as well be lol, hopelessly synthetic).

View attachment 185767

Interesting observation on the cache. I haven't really noticed a general decay in L3 performance up to 4600, just the weird 1/2 L3 read half the time. The L3 R/W/C numbers are so variable all the time I don't pay much attention to them until they pull a stunt like this.

Could it be a sleeping core related to power management? Windows Scheduler shuffling threads around?

 

[ViperXTR]

Power management? Maybe test it on high performance or balanced profile? Ryzen balance does not exist on Ryzen 5000 yes?

 

[dgianstefani]

Feel free to recommend me some settings, i have weird RAM - Hynix something

clean boot not much better

This is 2x32GB 3600 C18 sticks, you may be expecting too much from them

Show me a Zen timings screenshot and I'll see what I can do. But yes, 2x32gb is a problem.

 

[tabascosauz]

Could it be a sleeping core related to power management? Windows Scheduler shuffling threads around?

Power management? Maybe test it on high performance or balanced profile? Ryzen balance does not exist on Ryzen 5000 yes?

Tried all 4 different plans. Still doing it. Although I decided to move to the Windows HP plan instead, the Ryzen plans are doing some sus shit with boosting in CPU benchmarks, sometimes driving my multi core down low as a 2600X.

 

[biffzinker]

Ryzen balance does not exist on Ryzen 5000 yes?

Isn't Renoir Zen 2? No power plans are installed by the chipset drivers for Zen 3 Ryzen 5x00 series. Just the Windows 10 power plans.

Edit:

• Cezanne (Ryzen 5000 Mobile with Zen 3),
• Lucienne (Ryzen 5000 Mobile with Zen 2),
• Renoir (Ryzen 4000 Mobile, all Zen 2),
• Vermeer (Ryzen 5000 Desktop, all Zen 3),
• Matisse (Ryzen 3000 Desktop, all Zen 2)

 

[ViperXTR]

Isn't Renoir Zen 2? No power plans are installed by the chipset drivers for Zen 3 Ryzen 5x00 series. Just the Windows 10 power plans.

Yeah, Renoir is Zen 2, AMD mentioned that zen 3 no longer requires special power profile as windows can manage it fine now (maybe it had something to do with the CCX differences vs zen 2)

 

[biffzinker]

From over on Anandtech: Ryzen 9 5980HS
Core to Core Latency:

vs Zen 2

AMD Ryzen 9 5980HS Cezanne Review: Ryzen 5000 Mobile Tested

www.anandtech.com

 

[Radi_SVK]

That's... Really bad. I'm sorry to say. Your latency should be under 60 ideally, mine is 54. Your read and copy speeds should be above 60gb/s. L3 cache is also slower than it should beSorry

You not being really accurate here. Do you have the same Speed and timings? Doubt so... With timings of 18-22-22-42 @ 3600Mhz the 71ns latency is perfectly in line, even though less than perfect. Show me where did you see that with such timings he should have under 60ns, let alone 54ns just like yourself?
@Mussels increase your DRAM voltage to 1.4V and try to work your way down by one steps increments in timings and see whats the best you can get, I just like yourself have a 3600MHz RAM sticks that came with lukewarm timings of 20-23-23-43 @1.35V
with 1.4V I went 16-18-18-36 which gave me 68-69ns latency, which is not bad for a less pricey 3600MHz sticks and its probably best you can get.. Dont listen to @dgianstefani , you cant ever achieve 54ns or under 60ns with those RAM sticks and 68-69ns will be perfectly fine.

 

[Mussels]

Show me a Zen timings screenshot and I'll see what I can do. But yes, 2x32gb is a problem.

And yeah i knew i would never come close to a C14 kit, but thats the price i paid to get 64GB of ram in just two sticks


AS well as the kidneys i stole off that homeless child

 

[ViperXTR]

You not being really accurate here. Do you have the same Speed and timings? Doubt so... With timings of 18-22-22-42 @ 3600Mhz the 71ns latency is perfectly in line, even though less than perfect. Show me where did you see that with such timings he should have under 60ns, let alone 54ns just like yourself?
@Mussels increase your DRAM voltage to 1.4V and try to work your way down by one steps increments in timings and see whats the best you can get, I just like yourself have a 3600MHz RAM sticks that came with lukewarm timings of 20-23-23-43 @1.35V
with 1.4V I went 16-18-18-36 which gave me 68-69ns latency, which is not bad for a less pricey 3600MHz sticks and its probably best you can get.. Dont listen to @dgianstefani , you cant ever achieve 54ns or under 60ns with those RAM sticks and 68-69ns will be perfectly fine.

Hmm, i have the same exact timings for 2x16GB dual rank 3600Mhz (18-22-22-42) using XMP profile but i get around 66ns latency. have yet to tweak memory timings though, my ol 3300X doesnt want too much modification in IF clock, timings and frequency, hope my current AGESA 1.2.0.0 (beta) and 5800X memory controller i can get lucky

 

[tabascosauz]

@Mussels For MJR I don't think the primaries really scale so 18-22-22 might be the best you're gonna get. Can't remember if it's one of the ICs that scales negatively with voltage but I remember hearing about people doing 3733 and 3800 on pretty reasonable voltage, just with the 18-22-22 primaries. It is a 16Gb IC so it's a big boi, and it's not Rev.B.

That said, 550ns tRFC is a bit...it might not be CJR but it's still Hynix. Can it do something like 630 @ 3600? That'll bring you down to about 350ns.

I don't know how MJR handles RRDS/RRDL/FAW/WTRS/WTRL/WR so you might have to experiment a bit there.

 

[Radi_SVK]

Hmm, i have the same exact timings for 2x16GB dual rank 3600Mhz (18-22-22-42) using XMP profile but i get around 66ns latency. have yet to tweak memory timings though, my ol 3300X doesnt want too much modification in IF clock, timings and frequency, hope my current AGESA 1.2.0.0 (beta) and 5800X memory controller i can get lucky

I suppose many factors can also play a role, for example I will assume your RAMs came paired. I got initially 2 x 8GB,but after a while I have realised that I wanted to add more. So even though got exactly the same 2 x 8GB, unfortunately the second pair came with inferior chips, not speaking that I havent got the 4 x 8Gb paired.. your 66ns default is pretty nice, I guess you could bring it down to 60ns, assuming your timings are also default and you have some headroom.

 

[biffzinker]

My 3800X at 42.75/43 gets this for the L3 Cache.
View attachment 185746

The above is with Command Rate set to Auto.

With GDM on this what I get. Hopefully it doesn't bluescreen again.

That variation for the L3 Cache has me curious.

 

[8BitZ80]

Rebooted, set to 3800/1900 - no WHEA thanks to latest AGESA

Are you running AGESA 1.2.0.0 on the January beta bios?

My system is stable on AGESA 1.1.0.0 but it does post random WHEA errors in the logs.

 

[Mussels]

Are you running AGESA 1.2.0.0 on the January beta bios?

My system is stable on AGESA 1.1.0.0 but it does post random WHEA errors in the logs.

yes, the new 1.2.0.0 beta

Seems stable with no WHEAs here, and i got WHEAs when i first had the board and tried 3800/1900 out of the box

 

[ViperXTR]

I suppose many factors can also play a role, for example I will assume your RAMs came paired. I got initially 2 x 8GB,but after a while I have realised that I wanted to add more. So even though got exactly the same 2 x 8GB, unfortunately the second pair came with inferior chips, not speaking that I havent got the 4 x 8Gb paired.. your 66ns default is pretty nice, I guess you could bring it down to 60ns, assuming your timings are also default and you have some headroom.

Yes, its the cheapest 32GB i found so even though its CL18, i just snagged it last time. Ive already fiddled with the timings when i still have my 3300X, on 5800X, not yet. But my experience with 3300X was not very good tweaking the memory frequency and timings it did allow me to clock 1:1 for memory clock and IF clock stable. I assume the IMC on the 5800X would be a little better than the 3300X and can allow hogher frequency and IF clock plus lower timings?
Also, not that im also on dual rank on each stick so it that makes it also faster than most single rank setups

yes, the new 1.2.0.0 beta

Seems stable with no WHEAs here, and i got WHEAs when i first had the board and tried 3800/1900 out of the box

i can't even boot properly my 5800X without having UEFI settings crashing on me or showing me CPU or VGA errors upon boot before AGESA 1.2.0.0 beta, so i didn't get to experience one of those WHEAs

 

[Alyjen]

You can ignore this Aida L3 cache reading, it's wrong.
If you enable F-max enhancer in BIOS it'll go back to what it's supposed to be BUT your boost clocks & performance results will dive
If you do all core OC vs PBO it'll also be higher (usually)

I think it's related to how Aida boost cores during these tests or sth, for some motherboards it varies between BIOS versions, for others it doesn't. Like I said you can ignore this, it's not proving anything.

 

[Mussels]

You can ignore this Aida L3 cache reading, it's wrong.
If you enable F-max enhancer in BIOS it'll go back to what it's supposed to be BUT your boost clocks & performance results will dive
If you do all core OC vs PBO it'll also be higher (usually)

I think it's related to how Aida boost cores during these tests or sth, for some motherboards it varies between BIOS versions, for others it doesn't. Like I said you can ignore this, it's not proving anything.

Wait FMAX kills the boost? i've been flipping that on and off but not sure what it really does, vs what it says it does