Can somebody explain the CPU Offset vs Adaptive+Offset voltage implementations?

[toyo]

Hey there,

So, I am undervolting the 8700K. The board is a budget MSI Gaming Plus Z370, and it has a few voltage modes. I used Offset alone for a while at -0.090V, and all worked fine, but I thought I'd give Adaptive+Offset and the same negative offset to try it out.

To my surprise, it resulted in somewhat reduced power draw (125W vs 115W under LinpackX and prime95 SmallFFTs) at what looked to be the same voltage. It was also unstable in LinpackX, meanwhile -0.090V on Offset was stable. Another peculiar observation is that Adaptive+Offset is definitely affecting the VID too, not just the Vcore, while Offset is not. And another peculiarity: in HWinfo64, there are these new-ish Offset entries: IA Offset (these must be the cores), CLR (CBo/LLC/Ring) Offset and others. These 2 show the -0.080V Offset, both. Others don't, for example Uncore/SA Offset is still 0.000. Well, these offsets are shown ONLY in Adaptive+Offset. In the previous Offset mode I was using, they stay at 0.000, personally I thought they were broken or not fully implemented.

If you're tempted to ask "Why undervolt for a K SKU", well, I only had money for a 1070ti. To be fair, it's actually enough for what I play and at 1080p. The 8700K at 4.3 or 5GHz almost doesn't make any difference, if I still played games like WoW or something depending on a single core and with the GPU working at 50%, sure, but for now the 1070ti is the bottleneck in anything I play. I did kept the CPU OCed at 4.7-5GHz for a few months and it's OK, but also just useless lost electricity for what I do. As for why undervolt, boredom+temps are better and you can drop power draw in prime95 with about 25W or so vs stock.

Any ideas about these peculiarities? Thanks.

 

[Athlonite]

Any ideas about these peculiarities? Thanks.

sometimes the stuff they show in the BIOS hasn't been implemented ie: like a light switch on your wall if there's no light bulb at the other end no amount of playing with the switch will turn on the light

also with that tiny VRM probably not a good idea to be pushing it with high OC's

 

[toyo]

Any ideas about these peculiarities? Thanks.

sometimes the stuff they show in the BIOS hasn't been implemented ie: like a light switch on your wall if there's no light bulb at the other end no amount of playing with the switch will turn on the light

also with that tiny VRM probably not a good idea to be pushing it with high OC's

Well both Adaptive+Offset and Offset alone have a clear impact so they're implemented for sure. I'm just curious about why the first is resulting in less power draw than the second, and why HWinfo64 is only recognizing the offsets from Adaptive+Offset. The voltage decrease is still there for Offset mode alone, but HWinfo64 doesn't see it.

And also the VID, the Offset mode alone is not impacting it, Adaptive+Offset is. I'm sure there must be some reasons and explanations for all this.

As for the OC/VRM.

It wouldn't be a good idea, if what I do on the PC would actually result in the same power draw as stresstesting, which is really not even remotely close. For example, stresstesting with prime95 /w AVX at stock with uncapped PL is 135W. The VRM is kinda starting to not like it at 130-150W and gets to like 80-90C.
At the same time, prime95 without AVX is barely drawing 90W. A very demanding game on the CPU like AC Odyssey will spike to 70W rarely and mostly hover around 40-50W.
At 5GHz/1.35V, those spikes are 90W at most. Which is 5W below what a PL capped 8700K draws and realy safe. The VRMs barely get to 70s.
So it's all about realistic workloads for what you do, if you run stresstesting and AVX code on all threads all the time, of course, you'll benefit from something beefy like the Taichi or other 8 phase and above boards. But if you run more down to earth stuff like I do (games, browsing, work in Indesign/Photoshop/Illustartor), you won't even draw as much power (while at 5GHz) as a stock CPU that's under stresstesting.

At the same time, I feel kinda sad that my FX8350 in the other room has a 8+2 phase motherboard and I didn't have more money to invest in this new PC. But, for what I do, I can definitely run at 5GHz, didn't try higher yet and probably never will. I've ran HWinfo64 for whole days while working and playing at 5GHz and the VRM barely touched 70s in the hot summer and the CPU drew below 90W. This issue is with stresstesting the OC so you know it to be stable, which is kinda annoying as for 5GHz the 4 phase VRM can't really do AVX, so you're left with non-AVX testing, less stressful.

 

[Vario]

Offset adds or subtracts from the VID voltage through the entire frequency/vid table. Adaptive add or subtracts only in turbo mode. Adaptive would be useful for undervolting so you are able to stay stable at idle and lower clocks so you do not crash from too low of a voltage. Negative adaptive would reduce the turbo voltage but leave the idle voltage untouched.

 

[toyo]

Offset adds or subtracts from the VID voltage through the entire frequency/vid table. Adaptive add or subtracts only in turbo mode. Adaptive would be useful for undervolting so you are able to stay stable at idle and lower clocks so you do not crash from too low of a voltage. Negative adaptive would reduce the turbo voltage but leave the idle voltage untouched.

Thanks, this is what I've read online as well. Still puzzled about the reliably observable difference in power draw and stability from Offset and Adaptive+Offset and why a mode is recognized by HWinfo64 and VID readings, and another isn't. Gonna try to keep Adaptive+Offset since it uses less power for some reason.

 

[Vario]

May be that the motherboard alters the VID variable using adaptive whereas it adds a corrective value in offset, hence the different VID.

Something you might want to do is manually set your VCCIO and your VCCST to 1V, these are often overvolted unnecessarily by the motherboard when enabling XMP, make sure they aren't at a ludicrous setting.

Try experimenting with raising your uncore frequency, overclocking uncore (northbridge cache and memory controller) will improve latency noticeably.

 

[toyo]

May be that the motherboard alters the VID variable using adaptive whereas it adds a corrective value in offset, hence the different VID.

Something you might want to do is manually set your VCCIO and your VCCST to 1V, these are often overvolted unnecessarily by ram when enabling XMP, make sure they aren't at a ludicrous setting.

Try experimenting with raising your uncore frequency, overclocking uncore (northbridge cache and memory controller) will improve latency noticeably.

I assume you mean VCCSA.
Anyway, SA and IO are set to 1.15/1.10, and VDIMM at 1.2 (for XMP@3GHz). The RAM I have is weird, and not on the QVL. It took me almost a year to get it stable, not even joking. Must have been 100+ hours of p95/Blend, and that was after I lost ages using memtestx86 and other memory tests that were absolutely useless for my case. The default SA/IO settings are like 1.33/1.25 and VDIMM 1.35 as it's customary for XMP. Many months I would test on and off while playing with voltages, many months I thought the DIMMs actually require 1.35V to run at XMP, wrong, lost lots of time trying all kinds of combos of SA/IO, only to finally discover that I could decrease VDIMM, and after that, SA/IO were no longer that critical. The instability would result in occasional game crashes and very rare BSODs. Once I brought the VDIMM to 1.2V all went away.

Now it's finally fully stable, but boy did it take a lot of testing and effort. The last time I buy RAM not on QVL, that's for sure. The uncore is on default, so I think that means 4GHz for games and stresstesting. Could give it a shot at 4.3GHz see what happens. When I OCed to 5GHz I kept it at 4.4GHz.

 

[Vario]

Anyway, SA and IO are set to 1.15/1.10,

Looked them up just now,
VCCIO is the memory controller and shared cache
VCCSA is the system agent
VCCST is the sustain voltage for standby mode

The Intel Datasheet here https://www.intel.com/content/dam/w...heets/8th-gen-core-family-datasheet-vol-1.pdf has useful information regarding what these do and what the voltages should be.

You should be able to go lower on these two, but if you can't its not a problem just stay under 1.2V IMO. I am running 1.05V on both with high performance b die 3200 CL14 ram.

and VDIMM at 1.2 (for XMP@3GHz).

That is perfect.

The RAM I have is weird, and not on the QVL. It took me almost a year to get it stable, not even joking. Must have been 100+ hours of p95/Blend, and that was after I lost ages using memtestx86 and other memory tests that were absolutely useless for my case. The default SA/IO settings are like 1.33/1.25 and VDIMM 1.35 as it's customary for XMP. Many months I would test on and off while playing with voltages, many months I thought the DIMMs actually require 1.35V to run at XMP, wrong, lost lots of time trying all kinds of combos of SA/IO, only to finally discover that I could decrease VDIMM, and after that, SA/IO were no longer that critical. The instability would result in occasional game crashes and very rare BSODs. Once I brought the VDIMM to 1.2V all went away.

I am surprised that reducing the VDIMM voltage was the cure but glad to see that worked for you.

Now it's finally fully stable, but boy did it take a lot of testing and effort. The last time I buy RAM not on QVL, that's for sure. The uncore is on default, so I think that means 4GHz for games and stresstesting. Could give it a shot at 4.3GHz see what happens. When I OCed to 5GHz I kept it at 4.4GHz.

Try raising Uncore, it can run at same multiplier as core. 4.3 should be doable.

If I could I'd run mine higher using my bios, but for whatever irritating reason my bios will not raise it over 4.3. When I can overclock it inside Windows using Intel's Xtreme Tuning Utility it can go as high as my core multi.


I understand your motivations for undervolting, I am running my processor below what it can do as well. Like you, I was able to go over 5 GHz but could not perceive any difference in games beyond additional heat coming out of my PC case. I am running at 4.6 just to keep pace with the 9600K and not really for any other reason, my voltage is at my stock clock turbo voltage, 1.168 Vcore under load. With these settings, I never go over 55C in games.

 

[Vya Domus]

Do yourself a favor and put a fan over that tiny ass heatsink.

 

[toyo]

I am surprised that reducing the VDIMM voltage was the cure but glad to see that worked for you.

Yeah, this is why it took me so long to get the PC stable, because I simply didn't consider the VDIMM. Also tried to manually enter latencies, so that lost me time too. Sometimes I would think it was stable, but then some game crashed again, and Blend would have workers stop as late as 11-12 hrs in, which was super frustrating. To make things worse, the motherboard/BIOS would not boot at times with lower voltages on SA/IO and just give that message that "Overclock failed", and that further confused me and made me think that I need more voltage when I needed less. These DIMMs (ICs made by SpekTek by the way) are most definitely unstable at anything above 1.35V, maybe even lower. I guess that means that I will probably not be able to OC them much, the most I've seen them doing was 3500MHz, but I have some doubts that was stable, not to mention the performance gains are basically nonexistent.

As for the VRMs, I always wanted to put some fan to blow over them, when I stresstested for 5GHz I left the case open, and they got toasty anyway, 90C, which I assumed was probably 115C inside the VRMs, so borderline. They never throttled though. AVX was out of question at that frequency, and I don't think that the 4 phase can do prime95+AVX at more than 4.5GHz, the power consumption just goes above 150W and that's toast territory.

In the spring I want to buy better cooling, probably a Noctua NH-D15, and that would further complicate the situation, as I assume it will cover all that area. AIOs are expensive and I really don't wanna deal with potential leakage or pump malfunctioning, or those coldplate fins getting dirty etc.

Any advice on how should I go about this?

 

[Vario]

Yeah, this is why it took me so long to get the PC stable, because I simply didn't consider the VDIMM. Also tried to manually enter latencies, so that lost me time too. Sometimes I would think it was stable, but then some game crashed again, and Blend would have workers stop as late as 11-12 hrs in, which was super frustrating. To make things worse, the motherboard/BIOS would not boot at times with lower voltages on SA/IO and just give that message that "Overclock failed", and that further confused me and made me think that I need more voltage when I needed less. These DIMMs (ICs made by SpekTek by the way) are most definitely unstable at anything above 1.35V, maybe even lower. I guess that means that I will probably not be able to OC them much, the most I've seen them doing was 3500MHz, but I have some doubts that was stable, not to mention the performance gains are basically nonexistent.

As for the VRMs, I always wanted to put some fan to blow over them, when I stresstested for 5GHz I left the case open, and they got toasty anyway, 90C, which I assumed was probably 115C inside the VRMs, so borderline. They never throttled though. AVX was out of question at that frequency, and I don't think that the 4 phase can do prime95+AVX at more than 4.5GHz, the power consumption just goes above 150W and that's toast territory.

In the spring I want to buy better cooling, probably a Noctua NH-D15, and that would further complicate the situation, as I assume it will cover all that area. AIOs are expensive and I really don't wanna deal with potential leakage or pump malfunctioning, or those coldplate fins getting dirty etc.

Any advice on how should I go about this?

Edit:
NHC14S which blows air around the CPU socket.

 

[toyo]

Edit:
NHC14S which blows air around the CPU socket.

Yup, I've seen those coolers. Can't say I'm very excited about them, as the price is nearly as much as the top NHD15 which has 2 fans and should provide -3-5C at load compared with the top down solution, at least from what I've read. If I can think of something to put on the VRM with zip ties I will go for that. Might even go for some 240-280mm AIO, and some made up stuff for the VRMs. But that's for the spring.

Anyway, more weird stuff from this PC:

- can do 3300 MHz for RAM at the same 16-18-18 primary timings, same 1.2 VDIMM, 1.15 SA, 1.10 IO. 3400 at same timings seems impossible. Not even 1.37V is stable on VDIMM, even coupled with 1.35 SA and 1.3 IO. I know these are high, I just wanted to see if I can push the RAM to 3400, but looks like not for CL16, or maybe not at all. Weird SpekTek/ADATA RAM is weird.
- if I manually set the Ring/LLC to 4.5 for example, the motherboard will adjust voltage automatically like for a core OC, so I had to add more negative offset
- the Uncore seems to be limited to the same speed as the cores under stress; so for stock, 4.3GHz for everything, even though I tried setting it at 44 or 45. If I stop p95 and allow it to idle, it will reach the speed set in BIOS.
- with the -0.125V Adaptive+Offset, Ring to 44 (43 for p95), the Vcore under AVX SmallFFTs load is 1.112V, probably can further drop the voltage and remain stable. Worst power draw is 115-119W. VRMs start to get toasty at this sustained load, 75C after 2 hrs.

Other interesting trivia:
- this motherboard was (probably a wise choice) left with a default power limit set to the 95W TDP. IMO that's both fair and also respects the Intel recommendations and stock operation. Of course, that meant power throttling under stresstesting, 3.8-3.9GHz was common for SmallFFTs. After 5 BIOS versions or so, MSI suddenly removed the PL in the last BIOS. They're on Auto, but AIDA64 says they're set at 255W, both Short and Long Power limits. They used to be at 118/95. Of course, no changelog or anything, they just decided that they stopped caring.

 

[Vario]

Sometimes you can zip tie a fan to the rear case exhaust at a 90 degree angle so that it cools the VRM, though I am not sure whether it would be cooler pushing air down onto the VRM which is counter to the case airflow direction, or pulling air away from the VRM towards the exhaust.

You can look at the PH-TC14PE, I have bought two used ones for around $40 each off eBay in mint pretty much brand new condition so far, it is identical to NHD14/15 performance but fortunately they don't command the noctua premium for used heatsinks. The Scythe FUMA has been reviewed pretty well as a budget alternative to the NHD15 but with very close performance. If you go with a NHD15, make sure to get the NHD15S, it has better ram clearance, IMO. Only comes with one fan included with the S though it has two mounts just like the Non-S. The ram clearance is worth it I think.