Intel processors crashing Unreal engine games (and others)

[zir_blazer]

For context on what @zir_blazer is annoyed about: there are 3200 / 22-22-22 / 1.2V modules on the market, and they most likely lack an XMP profile. So the customer can pick one or the other, but never both.

Here is a better context.
EVERY DDR4 3200 MHz JEDEC module I found, was some low end module. Example:

Kingston Value KVR32N22D8/3 / DDR4 / 32 GiB / Single Profile
JEDEC 3200 MHz / 22-22-22 / 1.2V
No heatspreader. No XMP. No nothing. It is a completely dull module as the kind that you would find in a branded prebuild system.

Kingston Fury Beast KF432C16BB/32 / DDR4 / 32 GiB / Three profiles, JEDEC + 2 XMP
JEDEC 2400 MHz / 17-17-17 / 1.2V
XMP #1 3200 MHz / 16-20-20 / 1.35V
XMP #2 3000 MHz / 16-19-19 / 1.35V
Has heatspreader. Which is better than not having heatspreader, I suppose?

This same behavior you can find in other brands like GSkill. It is THAT hard for the above Fury Beast to have the JEDEC 3200 MHz profile from the Value series?

 

[lakedude]

...It is THAT hard for the above Fury Beast to have the JEDEC 3200 MHz profile from the Value series?

Wild guess that maybe the Value running 3200 22-22-22 at 1.2 V is slower but less "leaky".

I barely remember but I think I recall something about leaky transistors being faster in some cases?

Anyhow the Fury Beast seems to need 1.35 V to hit 3200 which (combined with maybe being leaky) might make it run hot, hence the heat spreaders. I think it is this increase in voltage from 1.2 to 1.35 that puts it out of spec and forces an XMP profile.

 

[zir_blazer]

Anyhow the Fury Beast seems to need 1.35 V to hit 3200 which (combined with maybe being leaky) might make it run hot, hence the heat spreaders. I think it is this increase in voltage from 1.2 to 1.35 that puts it out of spec and forces an XMP profile.

It needs 1.35V to do 3200 with 16-20-20. If it needed 1.35V to do 3200 / 22-22-22, it would be objetively WORSE than the Kingston Value module.
DRAM has some sort of 3-way Voltage / Frequency / Timmings curve. The Frequency and Timmings curve is part of the JEDEC standard and covers backwards compatibility so that you can always put a faster module than the system supports and let it downclock with tighter Timmings by letting BIOS using some formula to calculate other possible valid values (I recall reading that on Thaiphoon Burner site but they redesigned it and can't find the article). However, for this case you also have Voltage, and I don't recall ever seeing any sort of analysis that draw conclusions about whenever the binning quality of the DRAM chips of the low end modules like these (OEM-style vs low end enthusiast) could be equivalent from a binning perspective (Would running the Fury Beast at JEDEC 3200 be within specifications, or not?).
Point is, if you use default settings and don't enable XMP nor touch settings manually, the Kingston Value are the better modules cause they will work @ 3200 MHz whereas the Fury Beast will do 2400. The irony is that they were about the same price the last time I checked even thought, the heatspreader isn't free on the BoM and it wouldn't hurt on a less power hungry module, either.

Reelevant is that Alder Lake-S+ can do DDR4 3200 MHz across the full load spectrum (Up to 2 DPC / 2R per module for 4R total). I considered funny than a 4 * 32 GiB DDR4 @ 3200 MHz build could be potentially faster than a DDR5 that would downclock to 3600.

 

[lakedude]

It needs 1.35V to do 3200 with 16-20-20. If it needed 1.35V to do 3200 / 22-22-22, it would be objetively WORSE than the Kingston Value module.

I expect Fury needs 1.35V to do 3200 period. That it can do 16-20-20 makes it "better" than the Value which can only do 22-22-22.

Point is, if you use default settings and don't enable XMP nor touch settings manually, the Kingston Value are the better modules cause they will work @ 3200 MHz whereas the Fury Beast will do 2400.

For sure.

This picture explains what I was talking about regarding the leakage. It is likely that the Fury is faster but less efficient than the Value.

 

[igor_kavinski]

Which modules you have that default to 3600 MHz? If you have only two, it should be 4400.

It's ALIVE and very LOUD after last being used in MARCH-2023. Windows update annoyed me for about an hour and half.

Here's the module information:



EDIT: Ah, you are right. JEDEC #2 and #3 are only estimates for OC.

4600 is the max it will do so I started squeezing the timings and hit a wall also:

 

[igor_kavinski]

Updated from UEFI 4.01 to 11.12 and got this:



So it's finally working correctly now! God bless one year of UEFI updates

 

[bononos]

Gigabyte already pulled the bios. HardwareLUXX said it was using AC/DC LL of 1.7 mOhm instead of the recommended 1.1 mOhm.

.........

What does that mean? Not familiar with LL setting.

 

[zir_blazer]

What does that mean? Not familiar with LL setting.

AC Loadline is kind of a compensation mechanism for vDrop during Processor load. Per Intel specification, there is a minimum level of compensation (The higher the AC Loadline value, the more Voltage) required for different Processors, with most of the low end or low power desktop ones asking for up to 1.7 mOhms and anything above mid end up to 1.1 mOhms. The better the VR of your Motherboard, the lower the compensation you need (Which essentially means that a Processor with 1.1 mOhms max like a 13600K/14600K and above isn't intended for a really crappy board that would require more than that). The higher the AC Loadline, the more Voltage you provide, and if it is more than you actually need based on your board VRM, you're hurting power efficiency (And even performance, if power consumption push you over a Power Limit it hurts your turbo clocks). If it compensates for less than you need, you either have an unstable system as if you were undervolting below required voltages for proper operation, or some protection kicks in (I think it was CEP) that produces clock stretching and you have a stable system but with significantly reduced performance.
The actual AC Loadline of a motherboard should be tested with an Intel provided tool, the VRTT, and that value should be the default for that specific board - more and you're wasting power (Which is the safer choice), less and you're risking making it unstable. Problem here is that not even the Motherboard vendors themselves seems to have any idea about how to correctly configure AC Loadline as if they never actually tested it. This means that while you can know the BIOS provided values for AC/DC Loadlines, you don't know whenever these are the correct values as per VRTT or the vendor pulled them out of is ass.

 

[igor_kavinski]

Still can't go above DDR5-4600 CL28. So I went down!

 

[coercitiv]

Still can't go above DDR5-4600 CL28. So I went down!

Your memory kit is listed on the Z790 Sonic - Alder Lake QVL list as working @ 5600 MT/s with Bios version 1.17 from what I can see:

Are your DIMMs mounted in slots A2 and B2?

 

[igor_kavinski]

Are your DIMMs mounted in slots A2 and B2?

A1 and B1. Checking A2/B2 would need me to mess with the giant icesleet cooler. Guess I may try next week.

 

[DAPUNISHER]

This isn't a support thread. Start one if you need help. You are mucking this up to the point I don't know which thread I am reading.

 

[DrMrLordX]

On the topic of AC/DC LL on Z790:

If the mobo OEMs are pushing up their LLC settings to improve stability, couldn't end users maybe help stabilize things by increasing the cycle speed of the VRMs? I know I have that option on my AM4 board. Most high-end boards should have it. Basically you ramp up VRM performance at the cost of heating up the VRMs/making them less efficient, meaning they can react to voltage overshoot and dips more rapidly.

 

[zir_blazer]

On the topic of AC/DC LL on Z790:

If the mobo OEMs are pushing up their LLC settings to improve stability, couldn't end users maybe help stabilize things by increasing the cycle speed of the VRMs? I know I have that option on my AM4 board. Most high-end boards should have it. Basically you ramp up VRM performance at the cost of heating up the VRMs/making them less efficient, meaning they can react to voltage overshoot and dips more rapidly.

I was pondering exactly this, than the AC Loadline value may not be a 100% static value on a given Motherboard but instead could be affected by certain poorly explored VRM settings. I recall having seen an article at some point of someone benchmarking ripple differences based on VRM switching frequency settings (On MSI boards it should be under DigitALL Power on the BIOS OC section). If the net result is that the increased power of the VRM is compensated by a lower AC Loadline and you get a net lower power under load, it may even be a good idea to use that as default. However, go back to the previous point: Who THE HELL measures what the correct AC Loadline value should be for your board? You need THIS to measure it. No Intel VTRR here to do it myself, and even Motherboard vendors themselves seems to be quite lost about this entire topic.

I have been arguing about "what the default AC Loadline should be" since about two years ago, because the Coreboot port for the MSI PRO Z690-A uses the max AC_LL/DC_LL value depending on SKU (So either 110 or 170. Which I think that is the same thing that Gigabyte recently did, then backtracked and removed the BIOSes), and on my 12600K, using 170 vs MSI BIOS V1.3/V1.7 default of 80 caused a whole 23W extra power consumption, and lower turbo clock speeds/performance because it was hitting the Power Limits (Which on MSI BIOS didn't). But MSI themselves were inconsistent about this because another system that I build with a 12400 defaulted to 110. So the Dasharo developers were unsure, had no way to measure it themselves, and decided to go with the "safest" value, which they believed to be the max. They were not the only ones to do it that way, XMG also interpreted the max value as the default:

On Intel Core 12th Gen H-series, the default value is '230' and the BIOS allows any value between '1' and '230'. Entering the value '0' resets the value back to 'Automatic/Default', which is '230'.

And since Gigabyte recently screwed it up, too, is like if no one knows what the default is.

Where are the Intel field engineers when you need them?

 

[DrMrLordX]

@zir_blazer

Seems like it's time for some 13900k/14900k/s owners to do some stability experimentation with different VRM switching frequencies.

 

[zir_blazer]

Seems like it's time for some 13900k/14900k/s owners to do some stability experimentation with different VRM switching frequencies.

You need more than that. At this point you need something done very methodically, otherwise you will not be able to get a true apples to apples.

I found the VRM switching frequency voltage benchmark I recalled, it was made by ElmorLabs. Is not much but is better than zero.

Also worth mentioning is that Intel has some kind of specifications for the components of Motherboards VRMs. Reelevant to Alder Lake should be the Intel VR14 PWM and IMVP9 PWM (Intel Mobile Voltage Positioning) specifications, which are not public. The closest thing I found is this document from 2009 for the Xeon 5500 platform. Load Line is mentioned quite a lot. Due to lack of public data, I don't know what is the official word about how VRM Loadline (LLC) should interact with CPU AC/DC Loadlines, cause so far I always talk about those as if they were in a vaccum but I suppose that Intel PWM specifications should mention this.

 

[coercitiv]

I was pondering exactly this, than the AC Loadline value may not be a 100% static value on a given Motherboard but instead could be affected by certain poorly explored VRM settings. I recall having seen an article at some point of someone benchmarking ripple differences based on VRM switching frequency settings (On MSI boards it should be under DigitALL Power on the BIOS OC section).

There's this video from Buildzoid where he tests the effect of switching frequency on the MSI Z690 Pro-A.

But MSI themselves were inconsistent about this because another system that I build with a 12400 defaulted to 110.

Both systems were built with the same MSI Z690-A board? This might be explained by higher chip variance in the lower bins. Even with the aggressive Asus and Gigabyte LL settings only a small percentage of CPUs are unstable. It could also be that MSI decided not to bother optimizing for the lower bins and used Intel defaults (max values).

Seems like it's time for some 13900k/14900k/s owners to do some stability experimentation with different VRM switching frequencies.

Higher switching frequencies have no useful effect on many/most boards, as Buildzoid remarks around the 2 min mark in the video above. Owners of 13900k/14900k with no means to measure the impact on voltage regulation would be hunting in the dark for an improvement that may not be there or may also be influenced by their LLC setting. In other words they would have to completely forego AC Load Line and play around with Vcore, LLC and switching frequency in the hope they stumble upon something better.

Meanwhile, the same 13th/14th owners could gradually increase AC Load Line by small increments until the system is completely stable. That would be using Auto values for Vcore and stock LLC, and would be guaranteed to converge towards a stable configuration. In fact, the checklist for all the necessary settings is right in the last Intel email to the press:

Spoiler

Obviously this should be the job of mobo makers, not customers.

 

[DrMrLordX]

Higher switching frequencies have no useful effect on many/most boards, as Buildzoid remarks around the 2 min mark in the video above.

That's unfortunate.

Owners of 13900k/14900k with no means to measure the impact on voltage regulation would be hunting in the dark for an improvement that may not be there or may also be influenced by their LLC setting. In other words they would have to completely forego AC Load Line and play around with Vcore, LLC and switching frequency in the hope they stumble upon something better.

It shouldn't be that bad, especially since we're talking about out-of-the-box behavior of 13900k/s and 14900k/s CPUs on Z790 motherboards. So of course they'd be using default configurations, or better yet, default with only two changes: 1.1 mOhms AC/DC LL (as a precaution in case the board defaults to 1.7 mOhms) and different VRM switching settings. Sadly it would be a waste of time if what buildzoid observed was true for Z790 as well. And of course findings would only apply for an individual rig or MAYBE specific board + CPU combos.

Meanwhile, the same 13th/14th owners could gradually increase AC Load Line by small increments until the system is completely stable.

They could, but doesn't more AC LL encourage overshoot, which may be a source of the observed degradation?

 

[coercitiv]

It shouldn't be that bad, especially since we're talking about out-of-the-box behavior of 13900k/s and 14900k/s CPUs on Z790 motherboards. So of course they'd be using default configurations, or better yet, default with only two changes: 1.1 mOhms AC/DC LL (as a precaution in case the board defaults to 1.7 mOhms) and different VRM switching settings.

1.1 mOhms is the Intel maximum value for the S line with 65/125W "base power", mobo makers are supposed to test and establish the lower value that is appropriate for each of their boards. But for argument's sake, if the customer were to set 1.1mOhms, then there should be no need for further switching frequency tweaking, since the board should be perfectly stable (though probably inefficient). If they're not stable, I would argue the boards or the CPUs are not within spec.

They could, but doesn't more AC LL encourage overshoot, which may be a source of the observed degradation?

The way Asus and Gigabyte configure AC LL leads to voltage undershoot, and this is probably the main reason for stability issues even with brand new CPUs. The part of the configuration that encourages overshoot is the disabling of TVB features, because unlike what common sense would tell us, with TVB disabled the CPU apparently continues to boost to max clocks but without the frequency clamping that TVB enforces after a certain temperature. If there is any degradation out there, that is more likely to be the cause. Disabling TVB also leads to abnormally high idle voltages, since TVB is also responsible for lowering voltage at lower temps.

 

[DrMrLordX]

1.1 mOhms is the Intel maximum value for the S line with 65/125W "base power", mobo makers are supposed to test and establish the lower value that is appropriate for each of their boards. But for argument's sake, if the customer were to set 1.1mOhms, then there should be no need for further switching frequency tweaking, since the board should be perfectly stable (though probably inefficient). If they're not stable, I would argue the boards or the CPUs are not within spec.

Haven't some OEMs been setting AC/DC LL as high as 1.7 mOhms? Or did I misread that?

The way Asus and Gigabyte configure AC LL leads to voltage undershoot, and this is probably the main reason for stability issues even with brand new CPUs. The part of the configuration that encourages overshoot is the disabling of TVB features, because unlike what common sense would tell us, with TVB disabled the CPU apparently continues to boost to max clocks but without the frequency clamping that TVB enforces after a certain temperature. If there is any degradation out there, that is more likely to be the cause. Disabling TVB also leads to abnormally high idle voltages, since TVB is also responsible for lowering voltage at lower temps.

Okay that is kinda weird.

 

[coercitiv]

Haven't some OEMs been setting AC/DC LL as high as 1.7 mOhms? Or did I misread that?

Gigglebyte did, I pitty the folks with premium boards from this vendor, Intel or AMD.

 

[igor_kavinski]

Gigglebyte did

Hopefully I may never have to buy another mobo from them. ASROCK/MSI is my current preference, with ASUS only for certain cases (like their CSM boards).

 

[GaiaHunter]

AT explains that they presume an "average user" isn't going to get into BIOS settings at all so they use whichever profile requires zero user intervention.

XMP/EXPO require fiddling around in the BIOS whereas JEDEC should just come up by default.

I take this to mean they are not fiddling with any of settings like CAS, just using stock/JEDEC.

I'm not sure about the value of that stance.
If you are in the DIY market, turning xmp/expo is trivial compared to the assembly.

If you are buying a pre-built you should be looking for a different type of a review.

 

[lakedude]

I'm not sure about the value of that stance.

I agree that it is like an automobile review where they only drive a SuperCar the speed limit, however it does simplify the review process and make things "fair" in a way.

There are so many "knobs" to twist that could skew results. Just look at this thread, we don't even know where the line is between stock and overclocking anymore.

What would you propose?

 

[Thunder 57]

As I expected, Anandtech has yet to update the "review" of the 14900KS. They said they would run the benchmarks again over the weekend with Intel's new guidelines, but alas, nothing. I am starting to agree with @DAPUNISHER about our corporate overlords since the PC Gamer thing was posted the same day. And no one ever did explain why the (65W) 7900 was beating the 7950X in some of those benchmarks. Total clown show.