Ryzen 9000 series (Zen 5) memory discussion, OC

[tabascosauz]

Wait, so in order to reach higher FCLK you lower SoC voltage? What SoC voltage do you use for 2200 MHz?

What's maximum safe SoC voltage for Zen 5? 1.25V as with Zen 4?

I'm thinking about buying 2x24 GB G.Skill https://www.gskill.com/product/165/390/1693902423/F5-6400J3239F24GX2-TZ5NR.
Heading for 2133 MHz FCLK with Ryzen 9000 CPU on B650E board. Shame that G.Skill does not provide information on ranks, I need 2 dual rank modules.

Or would this next kit be of more use? (With 2:1 divider.)

F5-8000J4048F24GX2-TZ5K - Overview - G.SKILL International Enterprise Co., Ltd.

Trident Z5 series DDR5 memory is designed for ultra-high performance on DDR5 platforms. Featuring a sleek and streamlined aluminum heatspreader design, available in metallic silver or matte black, the Trident Z5 series DDR5 DRAM memory is the ideal choice to build a high-performance system.

www.gskill.com

AGESA now caps VSOC at 1.3V I think. So anything below that is "technically safe" but as usual lower is always better as long as stable. And add to that Fabric now liking lower VSOC now (may or may not manifest in your testing), really no reason to run above 1.2V for most people chasing an ordinary 6000 daily stable setup.

You could go for that 8000 1.35V kit if you wanted, but if you're just going to run XMP I don't know if the current wisdom applies for high freq DDR5. When OCers say that 8000 or 8200 finally pulls ahead of a tuned 6000CL30 setup, I think they really mean to say that they're also pushing 1.5V+ for like CL34 to get those results, not running 1.35V XMP CL40.

 

[IntelFanboy]

I think that some of the Zen4 that was exploding many months ago were supplied with 1.35+ up to 1.5v on SoC.

1.25-1.3v max I think, and I would go for under 1.2v if possible.

On my setup I must lower fclk to 2167 if I go above 1.2v soc. Others I have talked to have full 2200 stability at 1.3v doc so it depenfs on silicone lottery. 1.3v soc is the max supported, probably longterm safe, but I prefer below 1.2v soc myself.

1.25v is not safe, you're particularly degrading how far you can run stable in MCLK=UCLK mode. I degraded my chip that could do 6600 1:1 MCLK=UCLK at 1.28vsoc is 6 months to the point that 6600 is no longer possible even at the 1.3v soc cap. I have proof of the system being 24hr stable with zen timings and hours run on stress test. Now will spit errors instantly. Also came fom being able to post 6800 1:1 to not being able to post at all. Additionally silicon degradation is not something where its like "oh if you set the voltage to 1.49 you can run indefinitely with zero degradation but hey if you set it to 1.5 boom! the cpu gonna blow up"!!! Not how it works, if 1.3 is bad then 1.25 and 1.2 are equally as bad. The work around I found is to not run 1:1/ MCLK=UCLK mode. I just straight up run 6600/2200fclk and max out my bandwidth whichi s being limited by the fclk anyways, and end up with aida64 latency of 53ns. that's as low as it goes witht he current ram/ memory controllers we have. 50 will only happen with MCLK=UCLK mode. That 3ns penalty costs too much soc & vddio voltage imo, not worth the risk of degradation outside of quick benchmarking runs.

On my setup I must lower fclk to 2167 if I go above 1.2v soc. Others I have talked to have full 2200 stability at 1.3v doc so it depenfs on silicone lottery. 1.3v soc is the max supported, probably longterm safe, but I prefer below 1.2v soc myself.

I have not seen vsoc affect Zen 4/5 stability at all, unlike with zen 3 which actually had some effect. With zen 4/5 if you can do 2200fclk you will still do it even on 0.9 soc. Might have something to do with AMD disabling error reporting/ whea with zen 4/5. much harder to know you're unstable unless its freezing/ not posting. And due to the lack of error reporting you're unlikely to see a performance hit too. Way harder to tell.

 

[IntelFanboy]

I own one too (9950x) and honestly the IO die seems the same to me. Doubt I could push 2400 FCLK out of it. There may be refinements AGESA side and maybe even manufacturing improvements, but thats it.

I'm curious if anyone is unable to run gear down mode disabled like me. So far that is my biggest irritant.

Thank god its not just me who can't run GDM off on Zen 5.
I was going crazy for a minute because I thought that I was going to be able to straight up run Zen 4 settings with a few vddio/vsoc tweaks and call it a day, guess not haha.

After some tweaks I can post GDM off. Still the system is not confirmed "stable" only tested 1hr so far but I already know there are issues as I tried using the integrated igpu and running into instability issues, which shows instability with ram as thats the "vram" for igpus...

Another "BUG" I'm experiencing aside from that is that VDDIO after 1.2v won't post. Only way to post past 1.2v VDDIO is to set the XMP profile initially and work from there. (I'm aware the XMP profile does its on little changes such as tweaking impendences and what not) I can replicate those same "tweaks" and still can't post higher than 1.2 vddio. Back to what I was saying, so with the XMP profile I have 1.35 VDDIO/VDD/VDDQ to work with HOWEVER, If I want more VDDQ/VDD the system will now just refuse to boot. Something as simply as going from 1.35VDD to 1.36/1.34 and boom system wont post no matter what you give it 30min nothing/dead. So yeah overclocking kinda on pause until new AGESA patches this shit.

Using ASROCK motherboard for whatever it's worth, maybe its an ASROCK specific problem?

Looking forward to a response on the VDDQ/VDD/VDDIO voltage issues, are you also experiencing this?

As I was saying, if I don't initially set no XMP profile then I can set VDD/VDDQ to my liking (up to 2v if I really want to) but then I'm limited to 1.2 VDDIO (anymore wont post).
None of this were issues on Zen 4 but oh well.

 

[ir_cow]

1.25v is not safe, you're particularly degrading how far you can run stable in MCLK=UCLK mode. I degraded my chip that could do 6600 1:1 MCLK=UCLK at 1.28vsoc is 6 months to the point that 6600 is no longer possible even at the 1.3v soc cap. I have proof of the system being 24hr stable with zen timings and hours run on stress test. Now will spit errors instantly. Also came fom being able to post 6800 1:1 to not being able to post at all.

Not surprised by this considering 1.3 was super high for Zen 3. Not that I was ever able to run 6600:1 stable, 6800:1 never posted once for Zen4

I'm just going to assume that anything up to 1.3 SoC is safe until AMD says otherwise. If not RMA your CPU and try again

 

[Taraquin]

1.25v is not safe, you're particularly degrading how far you can run stable in MCLK=UCLK mode. I degraded my chip that could do 6600 1:1 MCLK=UCLK at 1.28vsoc is 6 months to the point that 6600 is no longer possible even at the 1.3v soc cap. I have proof of the system being 24hr stable with zen timings and hours run on stress test. Now will spit errors instantly. Also came fom being able to post 6800 1:1 to not being able to post at all. Additionally silicon degradation is not something where its like "oh if you set the voltage to 1.49 you can run indefinitely with zero degradation but hey if you set it to 1.5 boom! the cpu gonna blow up"!!! Not how it works, if 1.3 is bad then 1.25 and 1.2 are equally as bad. The work around I found is to not run 1:1/ MCLK=UCLK mode. I just straight up run 6600/2200fclk and max out my bandwidth whichi s being limited by the fclk anyways, and end up with aida64 latency of 53ns. that's as low as it goes witht he current ram/ memory controllers we have. 50 will only happen with MCLK=UCLK mode. That 3ns penalty costs too much soc & vddio voltage imo, not worth the risk of degradation outside of quick benchmarking runs.


I have not seen vsoc affect Zen 4/5 stability at all, unlike with zen 3 which actually had some effect. With zen 4/5 if you can do 2200fclk you will still do it even on 0.9 soc. Might have something to do with AMD disabling error reporting/ whea with zen 4/5. much harder to know you're unstable unless its freezing/ not posting. And due to the lack of error reporting you're unlikely to see a performance hit too. Way harder to tell.

1.2v will offcourse be better than 1.3v, but may also cause degredation. I have build several Zen 4 machines and the highest I gave set soc was 1.16v (worst imc possible, it needed 1.16v soc to run 5800 and 1.28v to run 6000). On my machine I run 1.08v at 6000 1:1. I will be surprised if that causes degradation. On setups I have build I set 1.2v soc as max limit.

I ran soc at 1.06v previously and got bsod + restart about once a week. After raising soc by 20mv system has been rock solid for 2 months. Too low soc can definately cause instability. On Zen 3 if caused performancedrop.

Fclk 2200 has no need for a floorlevel soc, but too high soc will cause throttling. I get performancedrop in linpack with fclk 2200 if soc is above 1.2v. below 1.2v 2200 fclk is faster than 2167, above 1.2v soc 2167 is suddenly a bit faster indicating some form of throttling.

 

[LittleBro]

1.25v is not safe, you're particularly degrading how far you can run stable in MCLK=UCLK mode. I degraded my chip that could do 6600 1:1 MCLK=UCLK at 1.28vsoc is 6 months to the point that 6600 is no longer possible even at the 1.3v soc cap. I have proof of the system being 24hr stable with zen timings and hours run on stress test. Now will spit errors instantly. Also came fom being able to post 6800 1:1 to not being able to post at all. Additionally silicon degradation is not something where its like "oh if you set the voltage to 1.49 you can run indefinitely with zero degradation but hey if you set it to 1.5 boom! the cpu gonna blow up"!!! Not how it works, if 1.3 is bad then 1.25 and 1.2 are equally as bad. The work around I found is to not run 1:1/ MCLK=UCLK mode. I just straight up run 6600/2200fclk and max out my bandwidth whichi s being limited by the fclk anyways, and end up with aida64 latency of 53ns. that's as low as it goes witht he current ram/ memory controllers we have. 50 will only happen with MCLK=UCLK mode. That 3ns penalty costs too much soc & vddio voltage imo, not worth the risk of degradation outside of quick benchmarking runs.

What RAM kit are you using? That's impressive latency for a DDR5 memory. I guess the tRFC/tREFI makes the most of the work at lowering latency, or not?
I have similar latency but on 5600X, DDR4-3800 @ 1.45V, 14-15-15-15-29, tRFC 264, Ol' Sammy's B-Die. I thought that such low latency on Zen 4/5 is nearly impossible.

Mostly, people have latencies such as:

And when I see this, I'm feeling cold sweat on my back ...

My 5600X runs at 1:1:1 (1900 MHz) with VSoC at 1.1250V. I wouldn't dare to go past 1.2000V for VSoC, especially not with X3D CPUs.

The IO die is the same for Zen 4 and 5. This means you can buy some that can only do 2066 stable, all the way to 2133. CPU lottery at play.

Are you planning on doing a Zen 5 RAM scaling test? How Zen 5 fares in few apps and games depending on RAM settings ...

 

[R-T-B]

Looking forward to a response on the VDDQ/VDD/VDDIO voltage issues, are you also experiencing this?

No, I didn't see that issue exactly, but did see odd gdm behavior. I'm on a gigabyte board though (X670 Elite AX)

 

[ir_cow]

Are you planning on doing a Zen 5 RAM scaling test? How Zen 5 fares in few apps and games depending on RAM settings ...

No short term plans to change my test setup to Zen5 this year or do any public scaling tests. Until the new Windows 11 build is public, I don't see the need.

 

[QyInvoLing]

No, zen5 didn't change its iod, so it remains the same as zen4.
Meanwhile, overclocking the IF bus could cause severe stability problem but without a crash, caused by IFbus's ECC feature, so just buy a pair of cheap memory and relax.

 

[Taraquin]

No, zen5 didn't change its iod, so it remains the same as zen4.
Meanwhile, overclocking the IF bus could cause severe stability problem but without a crash, caused by IFbus's ECC feature, so just buy a pair of cheap memory and relax.

If you check linpack you can easily spot if IF is running to high. Seems viritually all CPUs can run IF at 2133-2200. A few turds can`t do 2133 and some golden samples can do 2233.

 

[A Computer Guy]

The Zen experience

I'm so glad flashback has become a widespread somewhat standard feature on motherboards these days.

 

[ir_cow]

If you check linpack you can easily spot if IF is running to high. Seems viritually all CPUs can run IF at 2133-2200. A few turds can`t do 2133 and some golden samples can do 2233.

I guess mines a turd. 2133 isn't stable, never has been for me. Every AGESA update I check, same results. Does not pass y-cruncher 2.5b

 

[AVATARAT]

I guess mines a turd. 2133 isn't stable, never has been for me. Every AGESA update I check, same results. Does not pass y-cruncher 2.5b

Did you try to up VDD MISC or SOC_LLC a bit?

 

[ir_cow]

Did you try to up VDD MISC or SOC_LLC a bit?

Hmm now I think about this more, I haven't messed with VDD MISC much. It could be the key, though I'm a bit doubtful.

 

[Taraquin]

I guess mines a turd. 2133 isn't stable, never has been for me. Every AGESA update I check, same results. Does not pass y-cruncher 2.5b

What is you soc-voltage?

 

[ir_cow]

What is you soc-voltage?

It usually around 1.2v. but I tried 1.15 through 1.3v at 4800 JEDEC on different motherboards. I just accepted my CPU can't do 2133.

 

[Launcestonian]

I could not find any thread related to Zen 5 & memory support, so I'm creating one.

Are there any Zen 5 owners that run Ryzen 9000 series with Infinity Fabric at 2400 MHz? Or at more than 2000 MHz?

There have been numerous news posts about Zen 5 supporting IF freq. up to 2400 MHz. On the other hand, I am unable to find enough information on whether it is confirmed or not. Chips and Cheese wrote that Zen 5's IO die is the same as Zen 4's, thus IF runs at 2000 MHz and anything higher is not officially supported and is considered overclocking.

TPU wrote news post about 6400 MT/s being the sweet spot for Zen 5. This would require IF running at 2133 MHz. Even Wikipedia states 2400 MHz IF clock for Zen 5 (compared to Zen 4).

Zen 5 owners, what is your experience with Zen 5 so far? What clocks are you running your RAM and IF at?

2400 FCLK is wholly possible on 8000 series not 9000 as yet from what I've seen over the net.
Like any new tech its a constant work in progress for enthusiasts to tune well despite the commentary from jealous observers & Intel fanboys.
This is what I've done so far with AGESA 1.2.0.0a implemented by Asrock in bios 3.06 - Glad I could even do this all whilst customising the PBO at same time.

 

[Taraquin]

It usually around 1.2v. but I tried 1.15 through 1.3v at 4800 JEDEC on different motherboards. I just accepted my CPU can't do 2133.

Usually lower soc voltage makes high fclk possible. 2100 is very similar to 2133 anyways

 

[las]

6000 is still the sweet spot for Zen 5, yeah? I was seeing 6400 being mentioned at some point.

 

[Zach_01]

6000 is still the sweet spot for Zen 5, yeah? I was seeing 6400 being mentioned at some point.

6000 is said because it’s possible more frequent.
If UCLK:MCLK can do 3200 then it’s “better” depending also on the max FCLK possible and DRAM timings.

 

[las]

6000 is said because it’s possible more frequent.
If UCLK:MCLK can do 3200 then it’s “better” depending also on the max FCLK possible and DRAM timings.

However for 3D chips, it don't really matter I guess?

I remember seeing 7800X3D tested with diff memory speeds and it was a minor diff

 

[Zach_01]

However for 3D chips, it don't really matter I guess?

I remember seeing 7800X3D tested with diff memory speeds and it was a minor diff

True
3Ds are less affected because of the extra cache on package.
I guess it will depend on how much demanding the game is about memory.

 

[Taraquin]

However for 3D chips, it don't really matter I guess?

I remember seeing 7800X3D tested with diff memory speeds and it was a minor diff

It can be above 10% in certain games. In geekbench I get about 10% more after ramtuning with my 7800X3D. In general expect 5% avg vs 10-15% on non-3Ds.

 

[LittleBro]

7800X3D's memory scaling is more sensitive than 5800X3D's.
As Taraquin wrote, there are some documented cases where even mighty 7800X3D can get a nice perf. boost with faster RAM.
Or better to say, it's not thanks to a faster RAM but thanks to a tighter timings, 7800X3D does seem to benefit (a lot) in 0.1% and 1% lows at some games.

According to this video, the 7800X3D gets +30% fps boost in lows at Baldur's Gate 3:

Hardware Unboxed did ram scaling test some time ago and it shows some impact:

I'm aiming for 6400 MHz CL30-39-39-39-102 G.Skill kit with upcoming 9800X3D and will try to tighten it a bit.

 

[Taraquin]

7800X3D's memory scaling is more sensitive than 5800X3D's.
As Taraquin wrote, there are some documented cases where even mighty 7800X3D can get a nice perf. boost with faster RAM.
Or better to say, it's not thanks to a faster RAM but thanks to a tighter timings, 7800X3D does seem to benefit (a lot) in 0.1% and 1% lows at some games.

According to this video, the 7800X3D gets +30% fps boost in lows at Baldur's Gate 3:

Hardware Unboxed did ram scaling test some time ago and it shows some impact:

I'm aiming for 6400 MHz CL30-39-39-39-102 G.Skill kit with upcoming 9800X3D and will try to tighten it a bit.

Also worth mentioning: Most of the performanceimprovements is achieved by tuning REFI and RFC. Above 6000 7800X3D shows little scaling as there is limited bandwith on single ccds. The 79x0 with 2 ccds showed better scaling past 6000