Ryzen Owners Zen Garden

[The King]

I wouldn't trust it either, because it looks like it's in the market segment where the FETs are cheap as you know what. Increases the chance that you'll one day unexpectedly find the motherboard off and won't power on again.

Apparently Msi B650 entire range of boards are all 6 layers PCBs and maybe $hite when it comes to RAM support or OC.

 

[AusWolf]

Probably a smart idea. Skatterbencher has also been covering the newest AGESA 1003's Cstate boost limiter. Clearly the work on Raphael is still ongoing, chipset drivers implementing changes too.

20W > 11W is to be expected from going from a whopping 1.36V down to 1.02V VSOC

You mean 2400MHz UCLK or Fabric? 2400 Fabric would be surprisingly nice actually considering it's still the same design. Reminder that FCLK is no longer of primary importance (at least at EXPO/XMP speeds), it's just UCLK.

I might be confusing IF with the IMC. Are they on different clocks, anyway? It's strange that AM4 never confused me this much.

I mean this:

If I enable EXPO, the RAM runs at 6000 MHz, and the "Uncore" (whatever CPU-Z means by that) is at 1500.

Apparently Msi B650 entire range of boards are all 6 layers PCBs and maybe $hite when it comes to RAM support or OC.

Hmm... The big question, is it worth investing time into fiddling with it? Or would I be better off with a different motherboard? Or should I just run it at JEDEC spec and call it a day?

 

[The King]

Hmm... The big question, is it worth investing time into fiddling with it? Or would I be better off with a different motherboard? Or should I just run it at JEDEC spec and call it a day?

No experience when it come to AM5 or DDR5 so can't say. Maybe check this thread on this forum.

AMD DDR5 OC And 24/7 Daily Memory Stability Thread

https://www.amd.com/en/technologies/expo Soon to be released EXPO AMD optimized DDR5 memory. Likely as with Intel memory, A-die will be preferred. M-die is a close second. It has been said that like 3800-4000 1-1 the sweet spot for AMD 5000 series CPUs, 6000 1-1 should be the sweet spot...

www.overclock.net

 

[RJARRRPCGP]

Apparently Msi B650 entire range of boards are all 6 layers PCBs and maybe $hite when it comes to RAM support or OC.

Raphael now, is like Vermeer in 2020 and 2021.

 

[AusWolf]

Hmm... The big question, is it worth investing time into fiddling with it? Or would I be better off with a different motherboard? Or should I just run it at JEDEC spec and call it a day?

... or I could just go through a "turn on - wait a few seconds - turn off - turn on again and POST" cycle every time. What would you guys do?

 

[tabascosauz]

@AusWolf I'm not sure about average FCLK on Raphael but 2400 seems pretty impressive to me. If you know you can do 2400 FCLK stable, just set it manually from now on.

on AM4, I'm not sure if Uncore in CPU-Z refers to FCLK or UCLK. Functionally both were basically the same so we just treated it as FCLK.

In any case, you need to be looking at Zentimings anyway, not CPU-Z.

FCLK=Fabric
UCLK=memory controller
MCLK=memory speed

On AM4 there was just zero reason not to run 1:1:1 as a daily, which simplified things. On AM5, it's whatever:1:1 instead.

It's good to get Fabric up to at least 2000MHz, but most importantly you can't be running halved UCLK for good performance.

 

[AusWolf]

@AusWolf I'm not sure about average FCLK on Raphael but 2400 seems pretty impressive to me. If you know you can do 2400 FCLK stable, just set it manually from now on.

on AM4, I'm not sure if Uncore in CPU-Z refers to FCLK or UCLK. Functionally both were basically the same so we just treated it as FCLK.

In any case, you need to be looking at Zentimings anyway, not CPU-Z.

FCLK=Fabric
UCLK=memory controller
MCLK=memory speed

On AM4 there was just zero reason not to run 1:1:1 as a daily, which simplified things. On AM5, it's whatever:1:1 instead.

It's good to get Fabric up to at least 2000MHz, but most importantly you can't be running halved UCLK for good performance.

View attachment 271554

I'm not sure how impressive it is considering that it settles to 2400 at bone stock JEDEC settings and low voltages. I've seen a slide from AMD somewhere that "6000 MHz RAM at 1:1 with the IMC is the sweet spot". That 1:1 implies 3000 MHz FCLK, which is apparently not possible on my system - or maybe it is with some voltage tweaking, I don't know. I blame my motherboard, nonetheless.

The bigger question is how those numbers in the graph translate to real world performance. I mean, transfer speeds and latencies look nice in a diagram, but if my idle power consumption doubles with no more FPS in games, then what's the point?

This is my EXPO setting with everything else on Auto:

 

[tabascosauz]

I'm not sure how impressive it is considering that it settles to 2400 at bone stock JEDEC settings and low voltages. I've seen a slide from AMD somewhere that "6000 MHz RAM at 1:1 with the IMC is the sweet spot". That 1:1 implies 3000 MHz FCLK, which is apparently not possible on my system - or maybe it is with some voltage tweaking, I don't know. I blame my motherboard, nonetheless.

No, that's not what it implies. Like I said, 1:1 now only refers to UCLK:MCLK. You want to run 3000MHz UCLK; you can run whatever FCLK you wish.

• On AM4 it was 1:1:1 FCLK:UCLK:MCLK for good performance.
• On AM5 it is now auto:1:1 because Fabric, while not irrelevant, doesn't matter nearly as much as UCLK (like the slide above showed).

I'm 99% sure that 1:1:1 is straight up impossible at good DDR5 speeds (ie. 6000), the el cheapo chiplet IFOP design basically hasn't changed in any significant way since Ryzen 3000. Hence 2400 doesn't seem awful at all, most of the messaging implies that if you get to 2000MHz FCLK it'll be a non-issue. Auto:1:1 is a compromise between the needs of DDR5 and the reality that the chiplet design needs major changes if Fabric wants to keep up (ie. look at AMD's new fanout bridge tech on RDNA3).

But desyncing UCLK:MCLK is a different matter and analogous to Gear 2 on Intel (but from what I can tell, probably more along the lines of going Gear 2>Gear 4 since 12th/13th is optimized for Gear 2 on DDR5, while Raphael isn't).

As for game testing, if the game is purely GPU limited then you could run any CPU at base clock and 2133 JEDEC and not see much difference - doesn't mean it's a good idea, even if it saved a significant amount of SOC power.

You gotta install Zentimings my dude. CPU-Z doesn't provide any useful information.

 

[RJARRRPCGP]

What? 2400 for data rate with DDR5? If that's true, then it looks like a waste to go from socket AM4 and DDR4! (if 2400/2) (Reminding me of the early DDR2 days!)

OTOH, I would jive, if that's 2400*2. (4800)

 

[Mussels]

Damnit i forgot the voltage i used for my 5800x all core OC
My ITX system is having heat issues since i'm gaming on it while working on the custom loop in the main PC

(5800x + 1070ti can actually chew upto 350W, making it possible to use MORE wattage than my main rig gaming)

Was it 1.125v? 1.25v? dang it

 

[AusWolf]

No, that's not what it implies. Like I said, 1:1 now only refers to UCLK:MCLK. You want to run 3000MHz UCLK; you can run whatever FCLK you wish.

• On AM4 it was 1:1:1 FCLK:UCLK:MCLK for good performance.
• On AM5 it is now auto:1:1 because Fabric, while not irrelevant, doesn't matter nearly as much as UCLK (like the slide above showed).

I'm 99% sure that 1:1:1 is straight up impossible at good DDR5 speeds (ie. 6000), the el cheapo chiplet IFOP design basically hasn't changed in any significant way since Ryzen 3000. Hence 2400 doesn't seem awful at all, most of the messaging implies that if you get to 2000MHz FCLK it'll be a non-issue. Auto:1:1 is a compromise between the needs of DDR5 and the reality that the chiplet design needs major changes if Fabric wants to keep up (ie. look at AMD's new fanout bridge tech on RDNA3).

But desyncing UCLK:MCLK is a different matter and analogous to Gear 2 on Intel (but from what I can tell, probably more along the lines of going Gear 2>Gear 4 since 12th/13th is optimized for Gear 2 on DDR5, while Raphael isn't).

As for game testing, if the game is purely GPU limited then you could run any CPU at base clock and 2133 JEDEC and not see much difference - doesn't mean it's a good idea, even if it saved a significant amount of SOC power.

You gotta install Zentimings my dude. CPU-Z doesn't provide any useful information.

Ah I see. CPU-Z is unclear on what "uncore" means. Is it UCLK or FCLK? I'm wondering if HWINFO gives a clearer picture. I'll check in a moment. I'll also try manually setting FCLK to 2000 - maybe it doesn't like running at 1500 with high SoC voltage (which theoretically shouldn't be a problem, but who knows).

To be honest, I would gladly swallow the higher SoC voltage and +10 W idle power consumption if I didn't have the aforementioned cold boot lockup. Do you know any good 3D benchmark or game that is sensitive to RAM? I don't want to force my PC to do something it's obviously not happy to do just for +1% performance that I don't even see.

Edit: You're right, CPU-Z is garbage with this...

 

[tabascosauz]

@igralec84 maybe they improved things on Raphael but I tried Ryzen Master's CO testing feature on two 5700Gs and my 5900X. All were miserably nonsensical, recommendations were completely off the mark of what I knew was stable for each CPU (ie. would pass RM's test but literally nothing else since so unstable). On my 5700Gs I think I ran -10 to -15, run after run RM was calling for close to -30 all core.

Couldn't really tell why, from monitoring data it at least looked like RM was testing CO properly. Maybe the final output was just skewed

The usual tests still stand for CO, corecycler/occt/ycruncher etc., i havent seen anyone knowledgeable seriously consider testing with RM

Ah I see. CPU-Z is unclear on what "uncore" means. Is it UCLK or FCLK? I'm wondering if HWINFO gives a clearer picture. I'll check in a moment. I'll also try manually setting FCLK to 2000 - maybe it doesn't like running at 1500 with high SoC voltage (which theoretically shouldn't be a problem, but who knows).

To be honest, I would gladly swallow the higher SoC voltage and +10 W idle power consumption if I didn't have the aforementioned cold boot lockup. Do you know any good 3D benchmark or game that is sensitive to RAM? I don't want to force my PC to do something it's obviously not happy to do just for +1% performance that I don't even see.

Edit: You're right, CPU-Z is garbage with this...
View attachment 271623

Shadow of tomb raider is reasonably memory sensitive at like 720p or 1080p low I think. Can't remember exact settings.

I know it's not letting you 1:1 at 6000, but what if you ran 2400 FCLK manually (since you know it can do it), and set EXPO but brought UCLK and MCLK down to say 2800 (5600)?

I'm not familiar with DDR5 but don't flat 36s at 6000 suggest Samsung kit not Hynix? Think most OCers don't regard Samsung or Micron DDR5 very favourably at all, both seem harder to clock higher

 

[AusWolf]

Shadow of tomb raider is reasonably memory sensitive at like 720p or 1080p low I think. Can't remember exact settings.

I know it's not letting you 1:1 at 6000, but what if you ran 2400 FCLK manually (since you know it can do it), and set EXPO but brought UCLK and MCLK down to say 2800 (5600)?

I'm not familiar with DDR5 but don't flat 36s at 6000 suggest Samsung kit not Hynix? Think most OCers don't regard Samsung or Micron DDR5 very favourably at all, both seem harder to clock higher

If it only shows at 720p and/or low settings, then it doesn't matter - there's no point playing at low settings with a 6750 XT. Even if I did, having only a million FPS instead of a bazillion is of no significance.

Yes, it's Samsung, but that's not a problem for me. My attitude towards RAM has always been "set XMP and call it a day". All I want is for my PC not to lock up during cold boot when EXPO is enabled. Honestly, I only bought 6000 MHz RAM because "it's the sweet spot" according to AMD, and it wasn't a lot more expensive than any generic DDR5 kit.

I'll try the manual FCLK later, but I have my doubts that it'll help. With the Auto setting, it's currently doing 2000 with UCLK at 1500 and it still locks up during cold boot. Something else must be the issue. Maybe some of the voltages. Or maybe my motherboard really isn't the best at this.

 

[Chomiq]

@igralec84 maybe they improved things on Raphael but I tried Ryzen Master's CO testing feature on two 5700Gs and my 5900X. All were miserably nonsensical, recommendations were completely off the mark of what I knew was stable for each CPU (ie. would pass RM's test but literally nothing else since so unstable).

I gave it a shot once, took over an hour and a half (with two restarts) only to fail at 99% which then had me restart from scratch after reboot.

Instead I stick to -15 on silver/gold cores and -10 on everything else.

 

[Count von Schwalbe]

@AusWolf have you tried the UCLK at 3000 and FCLK at something lower, such as 1800 or 1600? It may be the fabric is the issue, as AFAIK it is the same as AM4. 2000 is high for AM4, at least.

 

[HD64G]

I gave it a shot once, took over an hour and a half (with two restarts) only to fail at 99% which then had me restart from scratch after reboot.

Instead I stick to -15 on silver/gold cores and -10 on everything else.

I preferred to just back off a bit from what RM co test suggested (-25 instead of -27 and -20 instead of -23 dependiong on the core) and has proven to be absolutely rock solid from there on.

 

[igralec84]

@igralec84 maybe they improved things on Raphael but I tried Ryzen Master's CO testing feature on two 5700Gs and my 5900X. All were miserably nonsensical, recommendations were completely off the mark of what I knew was stable for each CPU (ie. would pass RM's test but literally nothing else since so unstable). On my 5700Gs I think I ran -10 to -15, run after run RM was calling for close to -30 all core.

Couldn't really tell why, from monitoring data it at least looked like RM was testing CO properly. Maybe the final output was just skewed

The usual tests still stand for CO, corecycler/occt/ycruncher etc., i havent seen anyone knowledgeable seriously consider testing with RM

I agree, i don't think RM has improved that much, straight up giving -25 (or maybe it was even -30, can't remember but i thought there's no way that can be right) is a bit optimistic for all cores, that would probably mean the best ones are even more which i highly doubt.

I've settled for -15 now and +125 boost override and it's been stable so i'll stick with that. Maybe for benchmarks to improve my best scores i'll take some time and try the eCLK way with the raised multiplier and negative boost override, but for normal use and gaming, it's fine the way it is now. Maybe when i'm bored enough i'll investigate which CCX is better or maybe go core by core to see how good they are

 

[tabascosauz]

If it only shows at 720p and/or low settings, then it doesn't matter - there's no point playing at low settings with a 6750 XT. Even if I did, having only a million FPS instead of a bazillion is of no significance.

Yes, it's Samsung, but that's not a problem for me. My attitude towards RAM has always been "set XMP and call it a day". All I want is for my PC not to lock up during cold boot when EXPO is enabled. Honestly, I only bought 6000 MHz RAM because "it's the sweet spot" according to AMD, and it wasn't a lot more expensive than any generic DDR5 kit.

I'll try the manual FCLK later, but I have my doubts that it'll help. With the Auto setting, it's currently doing 2000 with UCLK at 1500 and it still locks up during cold boot. Something else must be the issue. Maybe some of the voltages. Or maybe my motherboard really isn't the best at this.

The Low settings are for the benchmark that ppl use for measuring memory/CPU performance. Not playing the actual game. Benchmark is free and included in the demo

6000 makes me a bit leery of both the board and the kit. Is it a dual rank kit? MSI only QVLs 2DIMM dual rank at 5800 for your board - on DDR4 their QVLs were conservative but DDR5 is a different animal, especially with low layer counts on MSI AM5.

I think TPU reviewed the same 6000CL36 kit you have. Not inspiring. The freq struggle bus seems common for Samsung and Micron which is why no one really takes them seriously compared to Hynix for DDR5. Like, I get that 12th gen IMC is mediocre like Raphael, but still, 0 freq headroom?

Corsair Vengeance RGB DDR5-6000 CL36 2x 16 GB Review

Corsair Vengeance RGB brings a competitive DDR5 kit, with an XMP 6000 MT/s profile and modest timings of 36-36-36-76. Designed to fit right into the Corsair ecosystem, it is time to take a closer look and see what all the fuss is about!

www.techpowerup.com

 

[AusWolf]

The Low settings are for the benchmark that ppl use for measuring memory/CPU performance. Not playing the actual game. Benchmark is free and included in the demo

6000 makes me a bit leery of both the board and the kit. Is it a dual rank kit? MSI only QVLs 2DIMM dual rank at 5800 for your board - on DDR4 their QVLs were conservative but DDR5 is a different animal, especially with low layer counts on MSI AM5.

I think TPU reviewed the same 6000CL36 kit you have. Not inspiring. The freq struggle bus seems common for Samsung and Micron which is why no one really takes them seriously compared to Hynix for DDR5. Like, I get that 12th gen IMC is mediocre like Raphael, but still, 0 freq headroom?

Corsair Vengeance RGB DDR5-6000 CL36 2x 16 GB Review

Corsair Vengeance RGB brings a competitive DDR5 kit, with an XMP 6000 MT/s profile and modest timings of 36-36-36-76. Designed to fit right into the Corsair ecosystem, it is time to take a closer look and see what all the fuss is about!

www.techpowerup.com

Yep, it appears to be the same kit, only that mine is EXPO, not XMP, and mine doesn't have RGB - which is why I chose it in the first place. Simple, plain design without the need for bloatware to turn the bling off seems to be a rarity these days. Performance-wise, as long as it runs at its rated EXPO speed, I'm fine - which it does if I start my PC basically two times (switch on - wait a few seconds - switch off - switch on again to properly POST), which is more of an inconvenience than anything, I guess. It's a single rank kit, so my board shouldn't have any problems with it (but it's MSi, so who knows).

I'll try the SoTR and RoTR benchmarks later, to see how much speed I'm losing at JEDEC compared to EXPO, but strictly with high graphics. I see no point in benchmarking with settings that I'm not gonna use.

 

[Ruru]

UPDATE

I may finally got rid of that audio latency problem! Did steps 4 and 5, haven't heard a single crack in audio after that *knocks on the table*

Fix High DPC Latency on Windows 11 / 10

DPC (Deferred Procedure Call) is a mechanism in Windows makes the system execute low-priority tasks later when the system is idle and high-priority tasks

thegeekpage.com

 

[Chomiq]

I preferred to just back off a bit from what RM co test suggested (-25 instead of -27 and -20 instead of -23 dependiong on the core) and has proven to be absolutely rock solid from there on.

-15 all core would end up in whea 19 reboots during low power states in my case. Can't imagine going above 20. Like I said, I don't even know what would RM suggest because it never finished.

 

[AusWolf]

I've just come across this article about the effect on RAM speed on gaming:

Fast RAM makes a difference in gaming

The Memory Performance Review - 6 Kits tested with 30 Games and Real World Benches Featuring Corsair DDR5 6400

babeltechreviews.com

I don't get it. When I look at the chart, I see ZERO difference between 6400 MHz, 6000 MHz, 5200 MHz and even 3600 MHz with a 12900K and 3080 Ti even at 1080p high settings, not to mention higher resolutions. Then the verdict says "We’d suggest that a gamer who prefers DDR5 choose a kit that is at least 6000MHz." Erm, WHY?!

I'm more and more on the conclusion that RAM speed doesn't matter at all for gaming, and I might as well jsut run at JEDEC 4800 MHz and get rid of the boot hangup issue, not to mention saving the extra 10 W on CPU idle power consumption.

 

[Ruru]

I've just come across this article about the effect on RAM speed on gaming:

Fast RAM makes a difference in gaming

The Memory Performance Review - 6 Kits tested with 30 Games and Real World Benches Featuring Corsair DDR5 6400

babeltechreviews.com

I don't get it. When I look at the chart, I see ZERO difference between 6400 MHz, 6000 MHz, 5200 MHz and even 3600 MHz with a 12900K and 3080 Ti even at 1080p high settings, not to mention higher resolutions. Then the verdict says "We’d suggest that a gamer who prefers DDR5 choose a kit that is at least 6000MHz." Erm, WHY?!

I'm more and more on the conclusion that RAM speed doesn't matter at all for gaming, and I might as well jsut run at JEDEC 4800 MHz and get rid of the boot hangup issue, not to mention saving the extra 10 W on CPU idle power consumption.

Intel platform isn't as sensitive to RAM speed as Ryzen. When I switched from 16GB @ 2666 to 32GB @ 3466 I saw a significant boost.

 

[AusWolf]

Intel platform isn't as sensitive to RAM speed as Ryzen. When I switched from 16GB @ 2666 to 32GB @ 3466 I saw a significant boost.

Apparently, Zen 4 isn't too sensitive, either:

DDR5 Ryzen 7 7700 DDR5 memory scaling review

You've likely just read our Ryzen 7000 reviews. and will have noticed that AMD calls DDR5 6000MHz /30CL a sweet spot. It, however might not be a sweet spot for your wallet, ergo we take a look at how... Performance - Gaming RTX 3090 - Assassins Creed Valhalla

www.guru3d.com

This calls for my own testing, I guess.

 

[tabascosauz]

Apparently, Zen 4 isn't too sensitive, either:

DDR5 Ryzen 7 7700 DDR5 memory scaling review

You've likely just read our Ryzen 7000 reviews. and will have noticed that AMD calls DDR5 6000MHz /30CL a sweet spot. It, however might not be a sweet spot for your wallet, ergo we take a look at how... Performance - Gaming RTX 3090 - Assassins Creed Valhalla

www.guru3d.com

This calls for my own testing, I guess.

Weird conclusion to come to since even in the limited selection of games they tested, there are instances where RAM makes a difference at all resolutions.

It's perfectly understandable that the games you play may not see much benefit, but blanket statements aren't very accurate.

Vcache generally seems to level the playing field, however, so RAM doesn't always make a huge difference