Ryzen Owners Zen Garden

[tabascosauz]

One can only hope... Imagine a 7800X3D fully unlocked except cache!

It still won't mean anything unless they overcome the thermal penalty that the 5800X3D suffers right now, but someone was saying that the design of the AM5 IHS suggests changes are afoot there. Maybe they'll design a new CCD around the Vcache instead of grinding the old one down to fit the Vcache.

If they can solve the voltage and temp problem, I'm thinking 7950X3D. The only thing Ryzen benefits from pure clocks is all-core productivity performance, pretty clear from the 5800X3D and Raphael Eco mode gaming results. The 6- and 8-cores are resigned to fight a losing battle against E-cores, but a 7950X3D that can have both cache and clocks has a clear advantage. I don't think AMD likes having 2CCD flagships that are beaten at gaming, and the 16-core will always be better poised to overclock (due to binning and its layout).

And barring other platform limitations, they could really just include 1 Vcache CCD and it'd do the job. Pretty obvious at this point that inter-CCD communication for gaming is still just as bad as it's always been due to lack of CCD-CCD link, so why bother with the other 8 cores?

If anyone has a 7900X/7950X and MW2, I'd be interested to see how many cores shader optimization can utilize now. Previously 1CCD CPUs treat it as an all-core workload, but 2CCDs and MW19 are limited to the contents of CCD0 only.

 

[freeagent]

One can only hope... Imagine a 7800X3D fully unlocked except cache!

The cache runs cool it’s everything else that gets warm

 

[Count von Schwalbe]

Maybe they'll design a new CCD around the Vcache instead of grinding the old one down to fit the Vcache.

Mebbe grinding the heatsink?

Pretty obvious at this point that inter-CCD communication for gaming is still just as bad as it's always been

It seems obvious that that is the reason that

include 1 Vcache CCD

Is not on the table. Or did you mean 1 CCD with V-cache?

If they can solve the voltage and temp problem, I'm thinking 7950X3D

Zen 4 X3D Limited to 8-Core and 6-Core, No Meteor Lake in 2023: Frosty Year Expected for CPU Market

A reliable source with CPU and platform leaks, ECSM_Official, made some new predictions about release timelines of upcoming desktop processors, and how 2023 could play out for Intel and AMD. 2022 is done, with no new desktop processor SKUs expected to launch from either brands. Intel is expected...

www.techpowerup.com

 

[tabascosauz]

Mebbe grinding the heatsink?

It seems obvious that that is the reason that

Is not on the table. Or did you mean 1 CCD with V-cache?

Zen 4 X3D Limited to 8-Core and 6-Core, No Meteor Lake in 2023: Frosty Year Expected for CPU Market

A reliable source with CPU and platform leaks, ECSM_Official, made some new predictions about release timelines of upcoming desktop processors, and how 2023 could play out for Intel and AMD. 2022 is done, with no new desktop processor SKUs expected to launch from either brands. Intel is expected...

www.techpowerup.com

Yeah, 1 Vcache CCD + 1 normal CCD, just entertaining the idea I guess. Would save costs since Vcache is completely wasted on the second CCD. 7600X3D and 7800X3D is the most logical but it doesn't get rid of the 1CCD tradeoff. Either you get gaming performance, or you get productivity performance - 7950X3D gets the best of both worlds.

The AM5 IHS remarks were about it being too thick, yes, speculating that they could thin the IHS on X3D products. But that assumes either that there's a new unique Vcache CCD, or that they've managed to stack Vcache with uniform z-height, without die thinning or silicon spacer. so it's just all fun speculation I guess

Some people think AM5's temps were caused by the thick IHS, just like the 11900K; personally I think it's just typical AMD being AMD, and pushing clocks/volts way way past the sweet spot of the process they're currently using. You've seen how cool and nice Zen 4 is at 4.7 base clock - at this moment that would be a great place for a X3D to be, since it's cache not clocks that are doing the heavy lifting in games.

 

[AusWolf]

Me neither, but the silly heatspreader isnt why

cold boot issues sound like the RAM or IMC settings are incorrect, and it then loads the JEDEC defaults and posts

You shut it down, it then tries to boot the bad settings, repeat. BIOS can show you say 3600 on the ram, but windows will show it's at 2133 when this sort of thing happens.

SoC is pretty likely to be all you need to change, but you'd wanna double check you've got XMP and no overclock/PBO settings anywhere (the various menus can conflict, you can do a per-CCX overclock, a curve undervolt and a static voltage and good luck knowing which one is active) - sometimes a clear to defaults is a good idea

PBO does not affect SOC voltage. If EXPO = XMP in function then setting that would automatically set VSOC as well, which is to be expected. 1.3V default VSOC........damn........reminds me of auto VCCSA going up to like 1.6V. Is it safe? Who knows?

After experimenting a little bit with the SoC voltage, the Auto option sets 1.2 V now for some reason. Still no luck with the lock-up during cold boot.

I've also read somewhere that over-tightening the CPU cooler can cause similar issues, so I loosened mine a little bit, but still no change.

I don’t know anything about AM5, sounds like it could use some vddg iod voltage?

VDDIO voltage is on Auto as well, at 1.36 V.

 

[Count von Schwalbe]

But that assumes either that there's a new unique Vcache CCD

That's what there was for AM4, due to the extra thickness. Thinning the AM5 IHS would prevent that issue, and probably improve thermals.

 

[Mussels]

Do you think PBO affects SoC voltage and I'd be good to go by just disabling PBO?

no PBO doesnt affect ram at all, and only indirectly alters CPU voltages (by limiting the clock speeds)

 

[outpt]

I thought that the long boot times was a bios fault and amd was going to fix these problem with new updates for am5 motherboards. Why is amd just now thinking about ihs problems? Surely this did not get passed all the engineers that work for them. I just can not believe that there is not a team working on vcache to help solve these problems since they have a good idea already.

 

[tabascosauz]

I thought that the long boot times was a bios fault and amd was going to fix these problem with new updates for am5 motherboards. Why is amd just now thinking about ihs problems? Surely this did not get passed all the engineers that work for them. I just can not believe that there is not a team working on vcache to help solve these problems since they have a good idea already.

The long boot times seem more to be DDR5-related, but AGESA isn't helping either. Since 2017 you have better odds betting on Half Life 3 release date than AMD proactively fixing AGESA did you really think that would change just for AM5?

I'm not sure what IHS problems you are referring to - Vcache delay is because TSMC delayed the approval of N5-on-N5 SoIC. 5800X3D's thermals are not really because of the IHS. Besides, no one knows what tricks they've got up their sleeves for N5 X3D. Maybe some creative packaging........or maybe no innovation at all.

 

[outpt]

From what I gather it is the thickness of the ihs they are using. A one core ccx already has heat problems and a thicker ihs would only compound the problems *. AMD in a effort to one up intel is now cranking up core clocks and voltage in a effort to stay ahead I guess. I can not believe they had no idea that raptor lake was stronger than than their upcoming cpus

 

[AusWolf]

no PBO doesnt affect ram at all, and only indirectly alters CPU voltages (by limiting the clock speeds)

Then all I can do is wait for an AGESA update, I guess.

Or maybe disable EXPO? Do you think that would have any effect on this problem? I've read somewhere that the performance drop in real life isn't that huge.

 

[Count von Schwalbe]

N5-on-N5 SoIC.

Forgive me if I am wrong, but doesn't cache scale badly with node? I would have thought N6...

 

[tabascosauz]

Forgive me if I am wrong, but doesn't cache scale badly with node? I would have thought N6...

Dunno. Some leaks had X3D as N6-on-N5, but that was from months ago, and TSMC tested a N6 on N5 application early on in 2021. But approved SoIC combos so far seems to be N7-N7 and N5-N5.

AMD's use doesn't even match the specs exactly; 5800X3D obviously used Gen 1 SoIC but with like double the bond pitch of what was advertised for SoIC 1.

 

[RJARRRPCGP]

They aren't exposed pipes, it has what looks like a nickel plate. Like the copper plate of the beloved early-212s, but what is probably nickel in its place, it doesn't look like the usual aluminum.

View attachment 269803

Yes! It came today! But, I didn't even open the box yet!

Installed, but will try a re-TIM sometime later.

 

[AusWolf]

I've just noticed something strange with my new system. When I cold start it, the CPU and RAM debug leds stay on, and it won't boot. If I turn it off and on again, everything is OK. What could cause this?

Another update on this: disabling EXPO makes my PC not only not lock up during cold boot, but it boots up almost instantly. I guess an AGESA update with improved RAM training is really in order.

Now the question is, how much detectable performance am I losing if I run my RAM at 4800 MHz CL40 instead of 6000 MHz CL36...

Edit: Other changes: SoC voltage is now 1.028 V and VDDIO is 1.1 V, and the memory controller runs at 2400 MHz (it ran at 1500 with 6000 MHz RAM). Do you guys think it would be worth manually setting these voltages with EXPO and a 1:1 RAM:IMC ratio? Or would it make it unstable?

 

[RJARRRPCGP]

Well, the temps at the default settings, are down to where Ryzen Master usually reports 81-83 C.

Update with two Noctuas, with push-pull configuration: Usually 81 C.

Note that this is in a warm room, IMHO.

 

[Mussels]

Another update on this: disabling EXPO makes my PC not only not lock up during cold boot, but it boots up almost instantly. I guess an AGESA update with improved RAM training is really in order.

Now the question is, how much detectable performance am I losing if I run my RAM at 4800 MHz CL40 instead of 6000 MHz CL36...

Edit: Other changes: SoC voltage is now 1.028 V and VDDIO is 1.1 V, and the memory controller runs at 2400 MHz (it ran at 1500 with 6000 MHz RAM). Do you guys think it would be worth manually setting these voltages with EXPO and a 1:1 RAM:IMC ratio? Or would it make it unstable?

In other words, your EXPO settings are too high for the sytem

turn EXPO on, but lower the RAM and IF down a few notches til you find where it's stable - that way you're getting the recommended voltages and timings

 

[Chomiq]

My co-workers gigabyte b550 board stopped functioning few weeks ago after multiple boot failures. It came back from RMA (twice), turns out that somehow Kingston's HyperX memory has failed critically, shorted and blew one of the fuses on mobo.

 

[tabascosauz]

Another update on this: disabling EXPO makes my PC not only not lock up during cold boot, but it boots up almost instantly. I guess an AGESA update with improved RAM training is really in order.

Now the question is, how much detectable performance am I losing if I run my RAM at 4800 MHz CL40 instead of 6000 MHz CL36...

Edit: Other changes: SoC voltage is now 1.028 V and VDDIO is 1.1 V, and the memory controller runs at 2400 MHz (it ran at 1500 with 6000 MHz RAM). Do you guys think it would be worth manually setting these voltages with EXPO and a 1:1 RAM:IMC ratio? Or would it make it unstable?

You sure your chip can't do 3000MHz UCLK if you manually set it? 1500 is 1:2 and not where you wanna be, 2400 is normal 1:1 for JEDEC 4800. Maybe bump VSOC a bit?

Fabric has been moved away but it sounds like VSOC is still responsible for the UMC itself - I have no idea what safe voltage range looks like for the N6 IO die but I don't think VDDIO (what is that? Is that MSI's name for Vmisc?) is helping you there.

AMD did say though that auto:1:1 @ 6000 is a "sweetspot". So essentially, good luck buddy, we ain't promising shit

Knowing all that, the long POST times make sense if it's just constantly failing 1:1 UCLK and eventually resorting to 1500MHz

 

[AusWolf]

In other words, your EXPO settings are too high for the sytem

turn EXPO on, but lower the RAM and IF down a few notches til you find where it's stable - that way you're getting the recommended voltages and timings

In other words, my motherboard is a PoS.

Which RAM and IF settings should I lower? The frequency? The RAM is designed to run at 6000 MHz, and the IF is already at 1:2 (1500 MHz) with the Auto setting.

Here are my voltages at JEDEC spec:
CPU NB/SoC: 1.028 V
DRAM: 1.1 V
DRAM VDDQ: 1.1 V
DRAM VPP: 1.8 V
CPU VDDIO: 1.1 V
CPU 1P8: 1.795 V

And here they are with EXPO on, all on auto:
CPU NB/SoC: 1.364 V
DRAM: 1.35 V
DRAM VDDQ: 1.35 V
DRAM VPP: 1.95 V
CPU VDDIO: 1.356 V
CPU 1P8: 1.8 V

I'm not yet familiar with how AM5 works, so I'm not sure what's right and wrong here.

You sure your chip can't do 3000MHz UCLK if you manually set it? 1500 is 1:2 and not where you wanna be, 2400 is normal 1:1 for JEDEC 4800. Maybe bump VSOC a bit?

Fabric has been moved away but it sounds like VSOC is still responsible for the UMC itself - I have no idea what safe voltage range looks like for the N6 IO die but I don't think VDDIO (what is that? Is that MSI's name for Vmisc?) is helping you there.

AMD did say though that auto:1:1 @ 6000 is a "sweetspot". So essentially, good luck buddy, we ain't promising shit

Knowing all that, the long POST times make sense if it's just constantly failing 1:1 UCLK and eventually resorting to 1500MHz

I've just tried the 1:1 setting with EXPO on and the system hung up during first boot (I waited 4+ minutes with no POST). I could get into the BIOS after a manual restart, but just barely.

Seeing the above, do you think I need more voltage somewhere?

 

[tabascosauz]

I'm not yet familiar with how AM5 works, so I'm not sure what's right and wrong here.

I've just tried the 1:1 setting with EXPO on and the system hung up during first boot (I waited 4+ minutes with no POST). I could get into the BIOS after a manual restart, but just barely.

Seeing the above, do you think I need more voltage somewhere?

I'm no expert either, I'm struggling just to figure out what the new rails do (turns out I was probably mistaken about Vcache getting its own rail). Raphael must not be very popular, have barely seen much OC discussion about it at all compared to 13th gen (particularly memory).

I'm just going off of skatterbencher:

Raphael Overclocking: What’s New - SkatterBencher

Raphael is finally here and just like with every new generation, a new CPU means new overclocking features.

skatterbencher.com

Looks like VSOC is chiefly responsible for mem controller. No idea what's a safe number and what isn't. Does the EXPO default 1.35V allow you to hit 3000MHz UCLK?

iirc Fabric you just run as high as you can run it, doesn't matter what ratio it is to UCLK. Same as AM4 (just that it was practically useless since there was no reason to unsync AM4)

The VDDIO doesn't look like Vmisc, the guide says VDDIO_MEM_S3. Having trouble understanding what it or VDDP functionally does yet, but it sounds like they would prevent you from booting 6000 if they were the issue (irrespective of what UCLK you had at 6000).

Correct me if I'm wrong, but on every other platform if your mobo is truly holding back memory speeds you wouldn't even post 6000 regardless of mem controller speed. It's not a high end board at 6400 qvl but it doesn't look like a board problem.

VDDQ and VPP are DDR5-specific voltages I think. As in, the sticks, not the memory controller.

 

[AusWolf]

I'm no expert either, I'm struggling just to figure out what the new rails do (turns out I was probably mistaken about Vcache getting its own rail). Raphael must not be very popular, have barely seen much OC discussion about it at all compared to 13th gen (particularly memory).

I'm just going off of skatterbencher:

Raphael Overclocking: What’s New - SkatterBencher

Raphael is finally here and just like with every new generation, a new CPU means new overclocking features.

skatterbencher.com

Looks like VSOC is chiefly responsible for mem controller. No idea what's a safe number and what isn't. Does the EXPO default 1.35V allow you to hit 3000MHz UCLK?

iirc Fabric you just run as high as you can run it, doesn't matter what ratio it is to UCLK. Same as AM4 (just that it was practically useless since there was no reason to unsync AM4)

The VDDIO doesn't look like Vmisc, the guide says VDDIO_MEM_S3. Having trouble understanding what it or VDDP functionally does yet, but it sounds like they would prevent you from booting 6000 if they were the issue (irrespective of what UCLK you had at 6000).

Correct me if I'm wrong, but on every other platform if your mobo is truly holding back memory speeds you wouldn't even post 6000 regardless of mem controller speed. It's not a high end board at 6400 qvl but it doesn't look like a board problem.

VDDQ and VPP are DDR5-specific voltages I think. As in, the sticks, not the memory controller.

I skimmed through the article, but it's more of a "what it does" kind of writing, and not a detailed guide of what is and isn't safe.

I think I'll just leave it at JEDEC 4800 MHz with IF at 2400 MHz (1:1) for now until more detailed guides come out. I don't really have the patience to manually tweak voltages myself.

Not to mention, running at JEDEC nearly halves my CPU's idle power consumption from 20 W to about 11-12 W.

I found this Guru3D article about the effect of different memory speeds with the 7700X. It looks like memory OC brings near zero difference with a 3090 at 1080p, so I should be OK with my 6750 XT.

 

[tabascosauz]

I think I'll just leave it at JEDEC 4800 MHz with IF at 2400 MHz (1:1) for now until more detailed guides come out. I don't really have the patience to manually tweak voltages myself.

Not to mention, running at JEDEC nearly halves my CPU's idle power consumption from 20 W to about 11-12 W.

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.

 

[RJARRRPCGP]

In other words, my motherboard is a PoS.

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.

 

[igralec84]

Has anyone tried the Ryzen Master Curve Optimizer test? I let it run for 1.5 hours yesterday, it restarted/crashed twice and them came out with -25 on all cores, which was stable maybe for one CB23 and Geekbench run. Since then lowered to -15 all cores and i think +125mhz boost override.

Been having some random crashes mostly right after logging on after upgrading from 7600X to 7950X, using the sam manual RAM settings (i have a non EXPO 5600 CL40 but set the primary timings from a Corsair 6000 CL30 since it's using the same Hynix M-die i think).

Not sure what it doesn't like but -10 and +100hz was stable for most Tuesday and yesterday before i started experimenting with CO and PBO again

I'm not getting any WHEA or bug check errors in event viewer, the crashes also don't seem to be logged as they're like hard resets, the only errors that seem to be happening are:

The SSGDIO service failed to start due to the following error:
A certificate was explicitly revoked by its issuer.
Event ID 7000

Doesn't seem like they were happening with the 7600X. I did have to reset my PIN at first 2 logins after fitting the 7950X, so not sure if there's some kind of fTPM or security issue or just over-optimistic PBO/CO. Or shit ASrock BIOS. Memtest does 3500% with the current settings so it's probably not the memory as i ran it like that for weeks before.