Ryzen 7000X3d is burning out.

[leadeater]

5 hours ago, Kisai said:

Both the CPU and the RAM are validated to run at the speeds advertised. If you only run at 2/3rds the speed advertised to be safe, you have to buy much faster memory to operate at JDEC speeds that the CPU supposedly supports. Note the part number even includes the 3200C16 speed in it.

 

So if you're not allowed to use XMP/EXPO, because it might cook the CPU, then who's fault is that?

Half right, the memory kit is validated to run at that but the CPUs are not. Intel and AMD only advertise maximum supported JEDEC memory speed at time of release and give support for XMP/EXPO at undefined speeds aka no guarantee.  

 

This is why we have situations at the end of the DDR4 era where not all Intel CPUs will run the fastest DDR4 possible to buy, you could choose a lower XMP speed in the BIOS but actually getting what was advertised on the memory box and specifications wasn't a guarantee in combination with a CPU.

 

Not all Intel 12th & 13th Gen CPUs will be able to run DDR4-5333/5200 which you can actually buy. It may be the CPU or CPU and motherboard combination but either way selecting 5333/5200 profile in the BIOS may result in the system failing to POST and reverting to base JEDEC.

 

As for fault that sort of depends, is the cause of the too high SoC voltage AGESA or in motherboard vendor implementations and customizations. If it was never safe to run over 1.3v then maybe AGESA should not allow it, but then there also needs to be support to disable that restriction for XOC but the default should not allow higher than 1.3. I think blame is probably a little shared.

 

DDR4-3200 is now a current JEDEC standard btw so you can run it without XMP/EXPO

 

 

Our newest AMD 7713 servers at work run DDR4-3200 ECC RDIMMs

 

Spoiler

 

[porina]

6 hours ago, Kisai said:

No, no no. XMP is "in spec" because when you buy 3200Mhz DDR4 modules, they are only operating at 3200Mhz in XMP mode.

This is the old "supported" thing again. From AMD's perspective, they only support up to JEDEC 3200 speed ram for DDR4 CPUs. XMP even at 3200 is not supported, which is not the same thing as saying it wont work. It can work, and often does, but AMD does not claim that.

 

What the mobo manufacturers and ram module manufacturers claim is separate to what AMD state.

 

1 hour ago, leadeater said:

DDR4-3200 is now a current JEDEC standard btw so you can run it without XMP/EXPO

Depending on the ram and mobo. It's a long term annoyance of mine that XMP modules typically only have a lower speed JEDEC profile. If you put it in a system that doesn't support XMP, you might only run at 2133 or 2400, even if the ram is XMP 4000. I can only speculate on the reasons why they don't do this. First is that the modules really can't meet JEDEC spec at higher speeds. As bonkers as this sounds, XMP is essentially ram overclocking and chips are likely ran outside of their base specification. Another possible reason is just how much space is there on the SPD chip? Maybe by the time they put in the XMP info they run out of space for more JEDEC profiles.

[leadeater]

8 hours ago, porina said:

Depending on the ram and mobo. It's a long term annoyance of mine that XMP modules typically only have a lower speed JEDEC profile. If you put it in a system that doesn't support XMP, you might only run at 2133 or 2400, even if the ram is XMP 4000.

Yea that annoys me too. It's really just down to the memory module choice, if you hunt really hard to the spec details you can probably find ones that use 3200 but I'd bet they would only be a lot newer revisions or products.

 

Kind of a really bad deal if you have to buy DDR4-5333 to get JEDEC 3200.

 

Edit:

@porina

I think this is what you'd want: KVR32N22D8/32 or AD4U320016G22-SGN or CT16G4DFRA32A etc. Not XMP though so ability to run them faster not as guaranteed as an XMP kit

[WallacEngineering]

This whole situation makes me wonder if Im better off forgetting AM5 all-together and instead snagging a Ryzen 5800X-3D and staying on AM4 to upgrade from my 3700X and not bottleneck my new RX 7900-XTX Red Devil.

 

Does anyone here know if Smart Access Memory would work between the 5800X-3D and the RX 7900-XTX despite them being 2 years apart in age?

[leadeater]

1 hour ago, WallacEngineering said:

Does anyone here know if Smart Access Memory would work between the 5800X-3D and the RX 7900-XTX despite them being 2 years apart in age?

Yes

 

1 hour ago, WallacEngineering said:

This whole situation makes me wonder if Im better off forgetting AM5 all-together and instead snagging a Ryzen 5800X-3D and staying on AM4 to upgrade from my 3700X and not bottleneck my new RX 7900-XTX Red Devil.

Would be a whole lot more value oriented this way and you'd be getting a large portion of the performance gain as well. Certainly the easiest option.

[Bitter]

I'm always a couple gens behind, mostly for cost but things like this keep me there.

[porina]

2 hours ago, leadeater said:

Kind of a really bad deal if you have to buy DDR4-5333 to get JEDEC 3200.

Maybe that's the memory module maker's thinking on it. Who would buy a high end XMP kit to run it at JEDEC 3200 or less? To me, it's more a convenience thing. My ram tends to go through various systems, and it would be nice to have higher JEDEC speeds on those systems that don't support XMP. And in those cases where people don't know about turning on XMP it'll give them a little bit more.

 

I kept a bunch of DDR4 just in case, as they're nice enough not to dispose of yet. G.Skill Ripjaws IV 3333, G.Skill Ripjaws V 3200 (2R), Corsair Vengeance Pro RGB 3200 (2R). All only run at JEDEC 2133 if the system doesn't support XMP. To be fair, JEDEC 3200 wasn't common when these Ripjaws kits were originally sold, but I'd take anything more than 2133.

[Bitter]

We should have a database of RAM and their JDEC info...

[WallacEngineering]

2 hours ago, leadeater said:

Yes

 

Would be a whole lot more value oriented...

 

Sweet, I might just go ahead and do that then, especially as I already have a Noctua D-15 with upgraded 2000 RPM fans and Trident Z NEO RAM. I just need storage upgrades as Im beginning to fill up with 3.25TB out of 4TB total filled.

 

Im thinking Samsung 980 Pro 2TB Boot Drive for Windows and core applications along with the most demanding games and then a "whatever" M.2 4TB to replace my Sata SSDs while also gaining capacity and then I can 7-Pass secure-wipe my old drives and sell them off to cover some of the cost.

[MageTank]

12 hours ago, porina said:

Depending on the ram and mobo. It's a long term annoyance of mine that XMP modules typically only have a lower speed JEDEC profile. If you put it in a system that doesn't support XMP, you might only run at 2133 or 2400, even if the ram is XMP 4000. I can only speculate on the reasons why they don't do this. First is that the modules really can't meet JEDEC spec at higher speeds. As bonkers as this sounds, XMP is essentially ram overclocking and chips are likely ran outside of their base specification. Another possible reason is just how much space is there on the SPD chip? Maybe by the time they put in the XMP info they run out of space for more JEDEC profiles.

It all comes down to cost. When you buy ICs from Samsung, Hynix, Micron, etc you have to pay more for certain speed grades which are essentially pre-binning. Alternatively, you can take an entire lot and perform the binning yourself. Those that do their own binning buy the base speed grades (likely 2133, 2400 or 2666 for DDR4) and perform the XMP binning themselves. This also means that every RAM "brand" has the opportunity to sell high performance kits and why I personally don't lend any credence to "X memory brand is bad, buy Y brand instead". At the end of the day, good RAM is based on IC quality and PCB revision, not the company slapping their name on it.

 

JEDEC binning is very different btw. They will NOT bin beyond 1.2V on DDR4. This means the DDR4 3200 JEDEC profile must confirm to 1.2V and will likely operate at CL22 (mid-grade as referenced in the chart provided by @leadeater). JEDEC 3200 binning does not guarantee any overclocking performance when pushing XMP bins. Just because a set of ICs are binned to JEDEC 3200 CL22 doesn't mean you can push 1.35V and drop CAS down to CL16. XMP bins are far more involved as you have to bin every single IC placed on a DIMM to operate at the increased speed, while JEDEC binning is often designed to look for less voltage leaking to operate within the 1.2V window.

 

There is plenty of space for several SPD profiles, you can add several profiles using Thaiphoon Burner yourself, but there is no real benefit for the manufacturers to add the extra effort into binning both low leakage for 1.2V AND IC quality at higher voltages for XMP performance. It just isn't cost effective to put the man hours into that.

 

Since we are on the subject of 3200 JEDEC, I'll share a fun little story with all of you. I was doing third party qualifications on an ASUS barebones system (GA35 desktop) which has an X570 STRIX motherboard in it. A solid motherboard for memory overclocking given the price. Advertised specifications listed 3200Mhz memory support, yet none of my 3200Mhz XMP profiles worked. When attempting to dial in a manual overclock to ensure automatic training wasn't busted, I could not find a VDIMM voltage rail to adjust. Turns out, they removed VDIMM voltage from the I2C bus and replaced it with an NFC reader because "gamers want to load saved games from an NFC drive on the go". This means the high end X570 STRIX motherboard was only compatible with memory operating at 1.2V, and the "3200Mhz" advertised memory speeds were for the JEDEC standard, lol.

[starsmine]

Me with a 7700X and an Asus Strix b650 mobo D:

 

[leadeater]

1 hour ago, starsmine said:

Me with a 7700X and an Asus Strix b650 mobo D:

Luckily you're "safer" with the 7700X since the SoC voltage is more within allowance. That OCP though... or absence of. Maybe you're not as bad since it's B series Strix and not X series hyper OC board.

[Vishera]

2 hours ago, starsmine said:

Me with a 7700X and an Asus Strix b650 mobo D:

 

ASUS - High SOC Voltage, Failure to implement proper OCP protection

GIGABYTE - BIOS so buggy that kills chips randomly (Therm Trip, VSOC Reset)

[AndreiArgeanu]

GN video is interesting. Yet to finish it but it looks like part of if not the problem lies with the motherboards?

Anyway this heavily reminds me of this old video where someone was pulling the cooler of a P4 and P3 where they thermal throttle and survive, then they show the AMD Athlon and Palomino which go up to 360C and burn up and die.

 

[ravenshrike]

22 minutes ago, AndreiArgeanu said:

GN video is interesting. Yet to finish it but it looks like part of if not the problem lies with the motherboards?

 

Both ASUS and apparently Biostar are shoving 1.4+ volts down the SOC at 'stock' settings when using EXPO even though the spec does not call for VSOC to change, probably in order to get a few percent increase of performance when using EXPO for bragging rights. Max recommended is 1.3, and preferably 1.25 or lower. Gigabyte... apparently figured out a brand new way to fry the CPU that AMD scratching their heads and scrambling to figure out what they did, and then make additional adjustments to AGESA to mitigate the possibility. AMDs failure is mostly not standing over the MB vendors' shoulder smacking a ruler over their knuckles every time they did something stupid. GN thinks they have more responsibility, but from AMD's point of view the mobo vendors were perfectly capable of handling x3D on the 5000 cpus so they probably didn't think it was an issue.

[CarlBar]

9 hours ago, ravenshrike said:

Both ASUS and apparently Biostar are shoving 1.4+ volts down the SOC at 'stock' settings when using EXPO even though the spec does not call for VSOC to change, probably in order to get a few percent increase of performance when using EXPO for bragging rights. Max recommended is 1.3, and preferably 1.25 or lower. Gigabyte... apparently figured out a brand new way to fry the CPU that AMD scratching their heads and scrambling to figure out what they did, and then make additional adjustments to AGESA to mitigate the possibility. AMDs failure is mostly not standing over the MB vendors' shoulder smacking a ruler over their knuckles every time they did something stupid. GN thinks they have more responsibility, but from AMD's point of view the mobo vendors were perfectly capable of handling x3D on the 5000 cpus so they probably didn't think it was an issue.

 

Someone in the comments of the video also mentioned this used to be a thing with intel and they, (Intel), clamped down on vendors doing it with 8th/9th gen. It's not like controlling this stuff should be a new concept. At the same time Asus and Biostar do deserve a lot of blame, going over 1.35 volets by .10mV or so is just MB vendors fudging it, but a full 50mV or more, thats just bad QA testing on the board vendor end.

[Shimejii]

On 4/22/2023 at 2:36 PM, Shimejii said:

Saw this yesterday, its an ASUS problem. Some Russian guy had pretty Identical issues with the a similar ASUS board, nearly in the exact same spot. Unless other board partners also start to have issues with the contacts burning up like this, sounds like its more so a ASUS and a trace problem. 

Seems like it wasnt a trace problem as much of a "How much Voltage? Yes" was the answer. Pretty surprised they got away with that in testing, why this was not flagged as a potential problem is just poor QC on Asus.

 

And  Gigabyte being buggy in their bios? No never! At least they have their safeguards working though

[xg32]

16 hours ago, Vishera said:

ASUS - High SOC Voltage, Failure to implement proper OCP protection

GIGABYTE - BIOS so buggy that kills chips randomly (Therm Trip, VSOC Reset)

the worry for me is that will 1.25v degrade the 7800x3d faster than other cpus, obviously 1.4v is just dumb. I still remember watching to touch 1.2v on my old chip.

[starsmine]

17 hours ago, leadeater said:

Luckily you're "safer" with the 7700X since the SoC voltage is more within allowance. That OCP though... or absence of. Maybe you're not as bad since it's B series Strix and not X series hyper OC board.

we'll see, 


1.3V, When I come back from my trip this summer Ill update the bios.
See if I can get away with dropping it to 1.25

[MageTank]

11 hours ago, starsmine said:

we'll see, 


1.3V, When I come back from my trip this summer Ill update the bios.
See if I can get away with dropping it to 1.25

I've pushed 6000Mhz C32 with 1.18V SOC. SOC scaling is a lot like VCCIO/SA on Intel where sometimes less is more. If you dial in the tertiary timings correctly, you can get away with significantly less voltage all around.

[HenrySalayne]

8 hours ago, MageTank said:

I've pushed 6000Mhz C32 with 1.18V SOC. SOC scaling is a lot like VCCIO/SA on Intel where sometimes less is more. If you dial in the tertiary timings correctly, you can get away with significantly less voltage all around.

The problem is stability. If you undervolt the cores, they will mostly fail with "easy" tasks. Cinebench is often running stable while the system is randomly crashing doing very little.

Especially with memory I wouldn't be sure if you even notice corrupted bits and other problems. This might result in sporadic system crashes, but it could also easily kill documents and other things you save. I would be careful without extensive testing.

[Bitter]

1 hour ago, HenrySalayne said:

The problem is stability. If you undervolt the cores, they will mostly fail with "easy" tasks. Cinebench is often running stable while the system is randomly crashing doing very little.

Especially with memory I wouldn't be sure if you even notice corrupted bits and other problems. This might result in sporadic system crashes, but it could also easily kill documents and other things you save. I would be careful without extensive testing.

I've experienced this a fought with this on my 2700X. Perfect stability under load, would randomly lock doing light tasks or fail to wake from sleep. I had to give up some peak clocks to get stable at light loads and transitional loads.

[MageTank]

15 hours ago, HenrySalayne said:

The problem is stability. If you undervolt the cores, they will mostly fail with "easy" tasks.

I am talking about memory though, not the cores. VSOC has little to no impact on core stability.

 

15 hours ago, HenrySalayne said:

Especially with memory I wouldn't be sure if you even notice corrupted bits and other problems.

I do, because I run primocache with write defer enabled. If I encounter even the slightest bit of memory stability, I lose my OS completely. You won't find a soul that walks this earth that can match my standards on memory stability, lol. Prime95 custom FFT for 24 hours, HCI Memtest 200% and 10 passes of Memtest86 with Linpack MKL tied in for CPU stability under heavy AVX.

 

My original point still stands. If you have a proper memory overclock with tertiary timings dialed in correctly, you can avoid running excessive SOC voltages. Also with AMD, it's not just SOC to watch out for. You need to be very careful with the VDDIO rails too. This isn't the old days where more voltage = more stability. Heat is a very real issue as density increases in these components and more voltage = more heat. More heat can cause just as much instability as less voltage, ask anyone that runs custom TREFI, lol.

[Quackers101]

GN video released

[Arokhantos]

Can we talk now about big brands etc trying to chicken out warranty any chance they can get even small user damage that can easily be repaired for small fee which totally reasonable ?

 

Damn i really miss EVGA, please make GPU's again even AMD gpu's and Intel GPU's

 

Its seriously frustrating you cant even do maintenance on your card while they try get away with stuff like this.

 

Asus is known to chicken out of warranty even if their board burned up your cpu.