Undervolting 265K on Z890 Tomahawk WIFI - BIOS bugs getting in the way?

[arnal160002e8]

Background:

My new PC has a 265K cooled by a Noctua NH-D15 (G1) on the Z890 Tomahawk WIFI in a Fractal Design Torrent case. While stress-testing at stock 1A90 BIOS settings (but with custom fan curves + Intel default power limits), temperatures in Cinebench R23 and Prime95 (Small FFTs with AVX/2) were too high for my liking (max. >85C in Cine, >100C in Prime). I reseated my cooler and re-applied the thermal paste, but to little or no avail. So, I decided to try undervolting the CPU.

To determine the minimum stable voltages for my CPU, I decided to use Intel Xtreme Tuning Utility (XTU), since that seemed like a more convenient way to quickly iterate on voltage offsets than rebooting and doing it from the BIOS each time.

My plan was to first identify the approximate minimum stable low (at idle) + minimum stable high (at full load) runtime voltages using XTU and moderate stress-testing, and then stop using XTU and instead lower my CPU Lite Load (default: 17 at "Auto") in BIOS to match my runtime voltages to those minima/maxima as closely as possible, fine-tuning with VCC Core Voltage Offset and CPU Core Voltage Offset if necessary. Then I would maximally stress-test those BIOS settings to determine the actual minimum voltages I could get away with.

That was the plan. Things did not go according to plan.

Problems:

Firstly, when applying undervolt offsets using XTU, I did actually see them being properly applied to the per-core VIDs (voltages) shown in HWInfo. And my temperatures did come down nicely with more aggressive offsets. Things seemed to be working!

But wait — VCore did not seem to be moving at all with the XTU Core Voltage Offset. And VCore was sky-high under load, often maxing out at >1.5V. And to make things worse, I was seeing wildly variable VCore values across reboots. Sometimes, VCore would max out at 1.490V during heavy load (Cinebench R23). Other times, the same load would max it out at just above 1.6V (!!!). This did not seem to correlate at all with my XTU undervolt offsets. I could have a -90mV offset and still have VCore top out at >1.55V. I later found out about this BIOS bug, which I assume is why VCore fluctuates so much between reboots with my VCC Core Voltage (VCore) set to "Auto" in BIOS. Furthermore, I read that allegedly VCore for Arrow Lake is the voltage supplied to the VRM before it is divied out as VID. So applying a per-core offset from XTU should not even be expected to alter VCore for Arrow Lake???

Anyway, not only did XTU not help me identify the minimum stable VCore, I now needed an updated BIOS in order to stand a chance of finding it. But at least I could still adjust CPU Lite Load to match my identified minimum stable VIDs... right?

Wrong. After identifying my presumed minimum stable core voltage offset, I went into BIOS to adjust CPU Lite Load down from 17. I had read (e.g. here) about users reporting that CPU Lite Load values below 13 just result in an effective value of 13 on the next boot, so I tried 12, and unfortunately confirmed what others had reported: 13 was displayed as the effective value when I next booted into BIOS. Same if I chose 9.

So I selected 13, thinking that surely it would at least make some difference to my voltages (VID / VCore) in Windows compared to the voltages I saw without an XTU undervolt offset applied and with the default CPU Lite Load of 17. But alas, any maximum per-core VID differences during Cinebench R23 were within the margin of error (~2mV for most cores) and VCore seemed unchanged or a little lower, although it was hard to tell due to the aforementioned inconsistent-across-reboots VCore bug. Temperatures were also unchanged. So my next question is: Does CPU Lite Load even do anything for Arrow Lake? If so, why can't it (ostensibly) be reduced below 13? Is this a BIOS bug? Update: CPU Lite Load does still seem to affect AC_LL/DC_LL. On my system, 13 translates to 80/120 while 17 is 120/120. Some claim that Intel recommends these numbers match. If that's true, a VCC Core Voltage offset seems like a better way of reining in VCore on Arrow Lake, but I'm not knowledgeable enough on the subject to say for sure.

I'd appreciate any advice on where to go from here for reducing my VIDs and VCore. I think I'm stuck until a BIOS comes out that fixes the inconsistent VCore issue.

 

[SmokeyGrayPoupon]

Using only BIOS, I am able to maintain CPU temps, running Cinebench r23, within an acceptable range by decreasing CPU Lite Load to a certain level, presently on 11. I can increase the level and see a corresponding increase in CPU temps as well as wattage in Cinebench r23. VCore stays consistent at 1.2-1.3. I also disabled IA CEP to restore performance in Cinebench r23 to 35850ish.
System specs in signature.

 

[arnal160002e8]

@SmokeyGrayPoupon Thanks for your input. With CPU Lite Load set to 11, is your BIOS displaying its effective value (in gray) as 11 also? Or as 13?

And may I ask how you ended up at 11? Did you test 13 at some point and observe an improvement in voltages/temperatures over that by dropping to 11?

Also, what monitoring software are you using to read VCore? HWInfo? Or some other? I think different monitoring programs may report different values for VCore (and not only because their sensor refresh intervals are not perfectly in sync), so an apples-to-apples comparison can potentially only be achieved if we both use the same program (and possibly the same version thereof).

Final question: In your signature, it says you are using BIOS 7E32v1A9. Is that 1A90 or 1A91? As you may know, there is a 1A91 beta BIOS available which was posted here to resolve the VCore inconsistency bug I referenced in the OP. (Unfortunately, that beta BIOS was unusable for me on the IGD.)

One thing I forgot to mention previously: my Cinebench R23(.200) score barely moved when I adjusted CPU Lite Load from 17 to 13. It was 35803 at 17 and 36025 at 13. I think the 222 point delta falls within the margin of error, as one of my XTU undervolt runs scored as "low" as 35572, which is 231 points less than the 35803 I scored at stock settings. Granted, it's possible that these scores were distorted by my having XTU and HWInfo open during some runs and not others.

 

[SmokeyGrayPoupon]

I actually started at default 17 and worked my way down, one by one, to a CPU temp that I was comfortable with, which is 78c-80c. Level 11 was the one. I will raise it one or two levels when the ambient temp turns cooler during winter. My present CPU cooler is able to keep the CPU temp ~ 50c-52c above ambient at 100% load at ~205-215 watts. I am using HWinfo64 to monitor my system sensors. I was not aware of a new Beta BIOS until your post. You are probably correct, the scores are distorted to a degree, HWinfo64 reduces Cinebench r23 scores by about 200-250 points. XTU, I have no idea, but I have to think it reduces the score as well.

 

[Nik Colch]

Hi!
I have the same configuration as you: Ultra 7 265K + MSI Z890 Tomahawk WIFI + Noctua NH-D15 (G1). Also have high CPU temperature during the stress test. I decided to order a new NH-D15 (G2) cooler, but it is still in delivery. I'll test it later to see how it works.

As for CPU undervolting. It seems that CPU Lite Load Control works a little differently here than on 13th and 14th generation motherboards. No matter how I tried to calibrate this setting, I still could not achieve the Vcore and voltages on the P-cores that I needed. And then I decided to completely ignore CPU Lite Load and focus on the core voltage offset, and this was able to bring better results.
So, here's what I did:

1. First, you need to set the "CPU Lite Load Control" value to "Intel Default".

2.0. Then you need to go to the "CPU Core Voltage Apply Mode" setting and set it to "All P-Cores & All E-Cores"
2.1. For the "CPU P-Core Voltage Mode" setting, set the value to "Offset Mode" or "Adaptive + Offset + Advanced Offset"
2.2. For the "CPU P-Core Voltage Offset Mode" setting, set the value to "-"
2.3. For the "CPU P-Core Voltage Offset" setting, you need to set the maximum value at which your CPU will work without crashes. For starters, you can set the value to 0.04, which will most likely work for most CPUs. After that, you need to run stability tests, and if there are no problems, gradually increase this value: 0.05, 0.06, 0.07, 0.08. etc. until the tests reveal instability. Then return to the stable value.

3.1. Now you also need to reduce the VCC Core value. So for "VCC Core Voltage Mode" set the value to "Offset Mode".
3.2. For "VCC Core Voltage Offset Mode" set the value to "-"
3.3. The minimum step by which you can reduce this value = 0.005, so start with it. And then again tests for stability, and then increase this value: 0.005, 0.010, 0.015, etc.
Attention! This setting allows you to reduce the Vcore voltage and at the same time affects the performance of the E-Cores. Since less current will be supplied to the E-Cores, there is a chance that you will not immediately notice the instability of the E-Cores. This may be revealed later when various errors occur: browser failure, game crashes, program errors, etc. I set this value to 0.015 and it turned out to be stable for me.
Also, I would like to point out that I don't think you should adjust the E-Core Voltage separately, since the VCC Core Voltage will also regulate the E-Cores, so keep "CPU E-Core Voltage Mode" set to "Auto", otherwise it will only add instability.

4. The next step is to set the CPU P-Core Voltage Limit. You can find this setting in the Voltage Related Controls section. I don't recommend changing any other settings in this section, just focus on the P-Core limit. The limit will depend on two factors: how good your cpu instance is, and what frequencies the P-Cores are running at. In my system, the P-Cores are currently set to 6 cores at 5.2 GHz and 2 cores at 5.4 GHz. For this combination of CPU frequencies, I set "CPU P-Core Voltage Limit" to 1.20. Other cores frequencies should have different limit.

That's all. You can also adjust the power limits if needed.

Finally, I will say that the problem with the BIOS actually exists, and the motherboard can randomly increase the voltage ignoring your BIOS settings. There's nothing we can do about it, we'll have to wait until MSI fixes this bug.

 

[arnal160002e8]

After flashing my MAG Tomahawk to the newest beta BIOS .1A92 (kindly supplied here by Svet), the inconsistent VCore across reboots is now fixed. This has enabled me to reliably gauge the impact of CPU Lite Load (CLL) at "Auto" (17) vs. 13 (the minimum permitted value on Arrow Lake). Here are my HWInfo64 read-outs after a cycle of Cinebench R23:

CPU Lite Load 17:                  CPU Lite Load 13:

Observed impacts of CPU Lite Load 17 -> 13 on my 265K:
- No statistically significant change in temps, clocks, or per-core voltages at idle or under load.
- Consistently reduced maximum VCore during a single cycle of Cinebench R23 from 1.480V to 1.448V, which is... a start.
- In BIOS itself, a reduction in VCC Core Voltage (VCore on Arrow Lake) from 1.436V to 1.410V.

 

[patsc158102e1]

@arnal160002e8
Nice that the voltages are now all stable. Now you could start optimizing your system and undervolting. Thanks for testing the .1A92 BIOS.

Every CPU/MB/RAM combo is, of course, individual, but I’ll still share the most important BIOS settings from my setup. Maybe you can find something useful.

 

[arnal160002e8]

Wow, that is a dizzying amount of altered settings...!

After skimming through the voltages section of SkatterBencher's 265K overclocking guide, I'm probably going to focus my optimizations on undervolting the upper extremes of the P- and E-cores' V/F curves, since those extreme points make the biggest difference to temperatures/voltages and should be the easiest to test for stability: just throw a high load at the relevant cores and see if they error.

Conversely, I'm not sure how I would reliably test for the stability of the lower points (i.e. ratios), so I doubt I'll bother with separate offsets for those and instead perhaps let their voltages be reduced by a modest blanket CPU Core Voltage offset that shifts the entire V/F curve down. (I would need to factor that blanket offset into the aforementioned V/F curve extremity point adjustments, of course.)

I'm also mentally preparing myself for having to lower the SOC SA Voltage once I enable XMP for my RAM, per this discussion...

 

[citay]

Observed impacts of CPU Lite Load 17 -> 13 on my 265K:
- No statistically significant change in temps, clocks, or per-core voltages at idle or under load.
- Consistently reduced maximum VCore during a single cycle of Cinebench R23 from 1.480V to 1.448V, which is... a start.
- In BIOS itself, a reduction in VCC Core Voltage (VCore on Arrow Lake) from 1.436V to 1.410V.

The reported power draw is also only minimally lower. Quite different to 9th...14th gen, where each mode lower would mean at least 5W less power draw under full load. Then again, we know we cannot really trust the sensor values of Arrow Lake fully, as mentioned here. This also makes undervolting a bit tricky, i think i would rely a lot on an external energy meter to really see how i make the most headway. I remember watching a Core Ultra review back when it came out, they changed some settings and it just wouldn't show up properly in the HWinfo sensors, despite there being a real change in how the CPU ran. But i cannot find that video anymore now. I wasn't this though, although it's related.

 

[arnal160002e8]

@patsc158102e1 I notice in your screenshots that you left P-core V/F curve point 7 at "Auto". While testing values, did you run into the issue I reported here?

 

[patsc158102e1]

@arnal160002e8
Yes, that's right. I had forgotten this "bug" again. As soon as values are entered at VFPoint 7, Turbo Boost 3.0 no longer works and the CPU only clocks up to 5400 MHz at most. That's why I leave VFPoint 7 on Auto. I enter the offset, which is also applied at VFPoint 7, in "CPU P-Core Voltage Offset." The application of the Advanced Offset Mode VFPoints is an additional offset that is applied to "CPU P-Core Voltage Offset."
It's good that you're bringing this up again. This has existed since the very first BIOS versions.

 

[arnal160002e8]

Thanks for confirming. I have submitted a formal support ticket referencing the thread I opened and have cited your comment as evidence that this bug is not specific to my setup.

 

[patsc158102e1]

Perhaps it’s a bug, or perhaps it isn’t. I had already adjusted to it for some time by leaving VFPoint 7 on Auto.

Likewise, it could also be platform-specific. The internal power supply with the Digital Linear Voltage Regulator(s) (DLVRs) isn't just designed for power savings and lower temperatures. There are probably also mechanisms at work that kick in when BIOS settings are too aggressive, so the CPU stays stable.

If you set a positive offset at VFPoint 7, Turbo Boost 3.0 won't break. Perhaps there are internal voltage thresholds that shouldn't be exceeded. And if they do, a kind of fail-safe mechanism kicks in and Turbo 3.0 is deactivated so that the voltage at the highest clock speed doesn't cause a crash.

I can't prove it, of course, but I've tried a lot of crazy BIOS settings, and the system has remained surprisingly stable, which I haven't experienced in previous PC builds. Windows blue screens would quickly pop up.

Thanks for your detailed explanation of when Turbo 3.0 will be inactive in your thread.

 

[arnal160002e8]

I had already adjusted to it for some time by leaving VFPoint 7 on Auto.

I would prefer to avoid setting a global CPU P-Core Voltage Offset to work around the VFP7 problem because:

(1) while an e.g. -100mV offset to the VFP7 reference voltage (1.35V according to SkatterBencher) would be "only" a ~7.4% reduction, the same offset to the VFP1 reference voltage (0.715V according to SkatterBencher) would be a relatively much higher ~14.0% reduction. And the difference between these percentages would only increase with a more aggressive offset.

(2) With such high undervolting at the low end of the V/F curve, stability-testing those low points would be a must (in my mind). But how would I reliably stability-test the frequency range below 5200 MHz? Stress-tests and benchmarks alike are typically designed to inflict maximum load on the CPU, which means running those would only test the highest frequencies my CPU is capable of at any given time.

Conversely, testing the 3 highest V/F points should be rather easy: with Enhanced Turbo disabled (or "Auto") in BIOS, the 265K's P-cores boost to 5200 MHz under an all-core load by default, so I can just run Linpack Xtreme, Prime95, and other multi-core torture tests to stability-test VFP5. Then, I can enable Enhanced Turbo from BIOS to allow the P-Cores to boost to 5400 MHz under an all-core load and repeat the earlier tests to stress-test VFP6. And if only this VFP7 undervolting problem didn't exist, I could then as a last step temporarily disable all P-cores except the two preferred cores to also stability-test 5500 MHz under a full (dual-core) load. And that would cover all three of the highest V/F points for my CPU.

I expect that stability-testing the frequency range below 5200 MHz would require down-clocking the CPU and multiple iterations of stress-testing at reduced maximum frequencies.

 

[arnal160002e8]

With the A92 BIOS, I am seeing odd behavior when a negative ring voltage offset is combined with ring V/F curve undervolting. I tried undervolting the 265K's 37x and 38x ring ratios by -220mV using two different methods:

Method (A): V/F Curve Undervolt
- Advanced Offset Mode For CPU Ring:
- 37x [-0.220]
- 38x [-0.220]
=> CPU Ring Voltage: 1.028V
=> VCC Core Voltage: 1.222V

Method (B): V/F Curve Undervolt + Offset
Advanced Offset Mode For CPU Ring:
- 37x [-0.170]
- 38x [-0.170]
CPU Ring Voltage Offset Mode [-]
CPU Ring Voltage Offset [0.050]
=> CPU Ring Voltage: 1.078V
=> VCC Core Voltage: 1.242V

Both methods should have yielded the same voltage at 38x, but didn't. Instead, it seems like the negative ring voltage offset of 50mV in method (B) was ignored by the motherboard. This was reflected in both the CPU Ring Voltage readout in BIOS (which was unexpectedly +50mV higher with the -50mV flat offset) and also the VCC Core Voltage readout in BIOS (which was unexpectedly +20mV higher with the -50mV ring voltage offset).

Can anyone confirm if this same behavior exists in the A93 BIOS?

 

[citay]

Just open a ticket with MSI, sounds like a "simple" BIOS bug.

 

[persiz154202d8]

Uninstall the Intel XTU, Clear CMOS and try this:
IA CEP - Disabled
CPU Lite Load - Mode 13 (0.800/1.200 mOhm AC/DC LL)
P Cores ratio - Auto
E Cores ratio - Auto
Enhanced turbo - Auto (Disabled)
Ring clock - Auto
D2D - 3200
NGU - 3200
VCC Core voltage - Offset -0.100v
P Cores voltage - Offset -0.100v
E Cores voltage - Offset -0.050v
Ring voltage - Offset -0.150v
VccSA - 1.070v
VnnAON - 0.770v
CPU IO - 1.250v

 

[patsc158102e1]

@arnal160002e8
I can confirm that the ring voltage behavior you described during undervolting also occurs in the A93 BIOS, as it did in previous BIOS versions.

It also affects the NGU voltage. A pure offset or a sole adjustment of the V/F curve works there too. If you combine both, the offset is ignored, which even had a negative impact on my VCC SA voltage, since the system agent, uncore (ring), and memory controller depend on it. But NGU voltage?

In any case, it's strange that the voltage listed by MSI in the BIOS as NGU voltage is displayed in HWINFO under Memory Subsystem Voltage (MemSS). The NGU Voltage item there remains unchanged. A mapping error in HWINFO? A reference to Asus, where their BIOS has separate voltages for NGU Voltage Offset and MemSS Voltage Offset. Or did MSI mistakenly name the Memory Subsystem Voltage as NGU Voltage? I have no idea.

 

[citay]

Ask in the HWinfo Forum, the HWinfo author Martin tends to respond there personally a lot.

 

[arnal160002e8]

The aforementioned ring voltage bug has been reported to MSI via web ticket.