Is undervolting dangerous?

Is undervolting a Skylake Mobile i7 dangerous in the sense that there could be CPU calculation errors?
Say its calculating 2+1 and it answers 4 there for corrupting data (?)

Oh and what Undervolts are you guys running on 6820HK/HQ because 98 deg. C isn't to good for my laptop I assume..

 

Dangerous? Just causes a WHEA error (not usually visible unless you monitor for them), an app crash or bsod/reboot. Just don't do anything important like install programs or let windows do updates before you're sure it's completely stable, where the calculation errors could mess things up (so have a system image or a full System Restore point or other backup)

The tolerances are pretty reliable. Run a 100% load benchmark (aida, XTU cpu stress test, TSBench, real bench encode, Intel burn test etc), slowly drop the core and cache offset until it crashes, dial it back 20mV from that minimum and you'll probably never have to touch it ever again.

 

As bennyg already said, it's unlikely you'll damage anything. The worst that can happen is that it will bluescreen. When you install something and it bluescreens it could happen that you need to reinstall whatever you tried to install later again.

 

Well, yes and no. When you are tuning the undervolt, finding the lowest stable undervolt - reducing the voltage to the CPU until it crashes and then backing off +5mV, you might need to back off a bit more.

Prime95 will start threads independent of each other and run them while checking results - if it finds a math error it will exit the thread with an error code. That's why running prime95 is a good test for undervolting - but you don't need to run it for hours, 15 minutes max is enough.

Then when you have a stable undervolt at 100% load testing, you do idle testing. When the laptop CPU is idle it will drop the voltage lower and lower, and if the undervolt is too low - it will crash or otherwise doing something odd - so you back off the undervolt another +5mV.

The combination of those 2 +5mV back off's should give you an idle/100% stable undervolt.

Note: I meant +5mV *increments*, keep tuning until stable. Some CPU's will be stable with +5mV once for 100% vs Idle, some will need more, increment and test.

If you are in a hurry I guess +20mV from 100% load undervolt might nail it right off, but you could also be giving up perfectly stable undervolt at +5mV or +10mV, if you don't have time to test right now go for +20mV and get back to having fun, and come back to see if you can get a little more undervolt when you have testing time, and patience.

That's why in the past I've said start at -100mV undervolt, and if you are stable and don't have time for more right now, stop there and you'll get that 10c temperature reduction at load, and then come back later for more undervolt testing when you have time. So far I've seen -150mV to -180mV 8th gen CPU undervolt range reported, so -100mV should be aok,.

Once you have that stable undervolt setting for idle/100%, you shouldn't get errors.

So, yes undervolted too far and you could get "math" errors, or crashes, or programs exiting oddly, but once you have a stable idle/100% undervolt, everything should run aok.

 

ehhhhh +5mv is maybe a little to much on edge, I'd say safe margin is like bennyg said +20mv. Sometimes the CPU will first bluescreen after 1 hour even if it ran stable with prime95 @ 15mins.

 

I'll update my post since that wasn't clear. I'm not saying stop there after backing off +5mV, I'm saying back off +5mV at a time until stable, and until you are not getting "math" errors - or other symptoms of being unstable.

If you started at -100mV and incremented -5mV or -10mV at a time until unstable, backing off in +5mV or +10mV increments should quickly get you to the 100% load stable value. I haven't seen a need to back off +20mV when incrementing in 5mV/10mV increments.

It's a long term thing too. I've found undervolt can be increased slowly over time, -1mV at a time, something to do when everything is running right and you want to induce instability - go back and try to undervolt some more. I took a -15mV CPU to -45mV stable in 18 months. Every little mV helps.

 

Seems a little time consuming for my taste, but yeah, you could do that.

 

Yeah, time consuming

Compared to re-pasting, unlocked BIOS OC'ing, and other extreme tuning - undervolting is a very brief exercise. Once you've done it a few times, it doesn't take long at all.

Start at -100mV test for stable, then I'd jump up to -180mV for 8th gen, -150mV 7th gen, and for 6th gen starting at -100mV I'd start incrementing -5mV/-10mV, test until unstable, then back off +5mv at a time until 100% load stable.

That shouldn't take long, maybe 15-30 minutes if you are taking your time. You could leisurely take longer, it's up to you.

Then back off +5mv or +10mV, or even +20mV to avoid idle BSOD, and you are done for the initial undervolt.

When you have time to play with tuning for a closer tune come back with smaller increments, but it's not really necessary if you are already running without thermal throttling or power throttling and are happy with temps and performance.

Take whatever pace or increment you want to keep it fun.

 

I would argue that isn't entirely true. As stated, sometimes the bluescreen will happen after some hours, it can take days to find the right calue.

 

You shouldn't have that happen if you are dropping back +20mV from 100% load stable undervolt, you should be stable right then and there.

For myself and most others I've helped / seen, +5mV / +10mV backoff from 100% load stable undervolt is enough for idle stable undervolt.

If you didn't do enough stability testing at the 100% load undervolt to make sure it is stable then I can see how it could BSOD later. Make sure you use a serious CPU load for 100% load stability testing. Do some gaming on that setting, run other apps, work on testing that 100% load undervolt before declaring it stable.

If you crash hours or days later at idle, then the 100% load undervolt wasn't really stable, so the additional backoff for idle undervolt won't be enough either.

Take your time going through the undervolting process, especially if this is new to you. It's not a race. Just because someone else can do it quickly due to experience doesn't mean you are expected to run through the process that quickly too.

The quicker you can determine an undervolt setting is unstable the sooner you can reduce the undervolt to test at a lower undervolt and reach a truly stable undervolt. You want to stress test it and make it fail, not find all the ways you can make it work.

I have seen some people declare a large undervolt over -220mV stable and then report they needed to drop back to around -180mV, so their initial CPU load testing wasn't adequate - don't use XTU Stress / Benchmark and declare stable, use Prime95 for final 100% load stability test.

If you declare -220mV stable due to inadequate CPU load testing, and then drop back +20mV to get idle stable, that's not going to be enough and you will be BSOD'ing down the road.

 

@hmscott @Danishblunt What P95 settings should I use? I have done a -99.6 mV Undervolt via Throttlestop on a 6700HQ

 

For a quick stress test select small fft, which hits the CPU hardest. Turn on your fan's 100% before starting. You shouldn't need to run it long, watch for individual restart of each thread - watch for failures where they exit and don't restart, that indicates a "math" error and too much undervolt - back off +5mV and try again.

For longer tests, balanced by how well your laptop cooling system can handle the heat, you can disable AVX / FMA extensions which unbalance the test - like when checking for core temperature differential you want AVX / FMA off - also for longer tests on computers without adequate cooling to sustain 100% load AVX / FMA disable it in the configuration file before running long sessions:

http://www.mersenneforum.org/showthread.php?t=21462

It says you don't have to exit, but I prefer to exit and start a clean session with all threads having started without AVX / FMA extensions.

Once you have the set up you like, with AVX / FMA enabled (default) or AVX / FMA disabled (edit config), you can quickly test undervolt setting with Prime95

It's really a matter of using your "math" to split the difference when going higher / lower undervolt to find the stable point.

In your case for the 6700 the undervolt won't typically have the same range as the 7th / 8th gen parts, although I have seen both high and low undervolting 6th generation CPU's too.

You are right at the sweet spot for undervolting near -100mV, which should give you about 10c decrease in 100% load temps.

One of the tricks I use to quickly find the stable undervolt is to jump too high (higher than I expect it to work) and BSOD. Like for the 6th gen, -120mV, crash and halve the difference between what minimum / maximum I hope for.

For a 6th gen CPU I'd hope for -75mV minimum, and -125mV maximum, so first I try -125mV, crash, then -100mV, then depending on the pass / fail at -100mV I go -110mV / -90mV, and then if -90mV passes, I go -95mV, if that passes I might stay there or shave -1mV/-2mV more - as you get better at the narrowing in on the stable undervolt, you can get "fancier" or finer grained.

Your -99.6 indicates you'e already narrowed it down a bit. If you have narrowed it down, and you think that's "it", then Prime95 for 15 minutes should confirm it for 100% load. Then I'd do the idle test - others may decided to drop back +5mV/+10mV and secure themselves against an Idle BSOD, but I take the time to find that point - usually I have more than 1 computer so it's no loss to my productivity to take my time at this.

If I was doing this for someone else "at their place", I might take the shortcut route - +10mV and set up XTU (TS) profiles for default settings and -90mV so if they are unstable after longer use at -90mV they can disable XTU (TS) or select the default profile and keep going until I can help them with figuring out of if the crash is undervolt or something else.

Yes, Windows crashes for many reasons and coincidental crashes while tuning are fun twists on otherwise simple routine.

 

@hmscott I ran a 3hr Prime95 Small FTT test default config, 8 threads no errors, is that good enough? Should I go lower because 99.6 was just my starting point...

 

Yeah, it's hard to tell isn't it

For a 6700 that's probably right at the point of best guess first time, you could just stay with that and go back to enjoying your laptop gaming, work, etc and not spend any more time on it.

There is a point of diminishing returns, and you are right at the start of that, so considering what the benefits are vs time invested, as long as you aren't thermal throttling during use there really isn't much more to gain on the face of it.

But, if you are chasing benchmark numbers, there's always room to eek a little more benefit out of every tuning. It's possible your 6700 is good for -120mV or even more for just that much more increase in power budget / thermal reduction, but it won't really be enough to notice, except again for benchmarking.

There is another "trick" TS does that XTU doesn't - XTU clocks 6th gen and new CPU's core / cache undervolt together - no more undervolting the cache further than the core to gain more power budget / reduce thermal load.

TS allows you to set 2 different values in the registers that will be ignored for actual differential core / cache undervolt settings, but other calculations seem to take those differing values into account and "trick" the CPU into allowing for more power to the CPU increasing benchmark "numbers" a tad further if you set the cache undervolt to -100mV more than the core undervolt.

@unclewebb 's description is here:
http://forum.notebookreview.com/thr...-7700hq-undervolt.800689/page-4#post-10757542

I only mention it as you use TS. It's only a few points in benchmarks, and you won't notice it day to day use, so to me it's not worth the time spent doing further tweaking, but if you are into benchmarks it might be useful to your top scores.

So, TL;DR - leave it there for now and run for a while, use hwinfo64 logging to capture sessions of gaming and apps that push the limits of the CPU to see what your real second to second temps are, not just the single Max peak value.

You can always come back to undervolting, but getting stuck in a seemingly never ending rathole eeking out another -1mV is fun for some but soul crushing for others, give yourself a break and enjoy what you've got.

 

@hmscott I have been running my CPU for 9 hours @ 98 deg. C, I am afraid I just shortened its life a whole lot. I also went to -125 mV and so far 3 tests passed P95 in about 10 minutes although CPU temp is still high AF and outside Intel specifications.

 

Nah, the CPU would have shut itself down at 100c. The CPU's these days are great at protecting themselves.

Also, why weren't you watching temperatures?

Did you have 100% fans on, or is that autofan? You probably also didn't follow my recommendation to disable AVX / FMA before running Prime95 to reduce power / temp load due to those instructions being so power hungry. Disable their use in config before running Prime95, see my other post in this thread.

Why did you run Prime95 so long? Didn't you see my 15 minute maximum run time recommendation?

There is no need to run long 100% load tests, it's mostly the idle undervolt failures that fail after hours or days, and if you exit all apps, shutdown all services, and stop using the laptop(!) allowing the laptop to quickly quiesce to lowest power usage - allowing CPU voltage to quickly drop to it's lowest, you won't need to wait long for that test either.

So, stop running Prime 95 for hours on end, it's not necessary, and probably harmful to your ears listening to the 100% fans for so long.

 

Okay I've put the CpuSupportsAVX=0 flag in P95's config and the laptop is maxxed out in terms of fans and I think my -125mV is stable and 2 minutes in, its holding up well...

EDIT: No its not, I got a P95 rounding error a few minutes ago.
EDIT2: went to -100.6 mV, should be stable. P95ing it now.

 

You can easily undervolt to -150. The CPU can usually handle around 165 - 180mv.

 

@Danishblunt P95 was throwing errors at 125 mV so I dialed it up a notch to 100.6 and did an idle test for about 30 mins, all stable. Is 30 minutes of Idle testing enough @hmscott

 

Today's 8th gen seem to be hovering around that range - about -170mV, but the 6th gen CPU's were much lower on average, around -125mV. Haswell and Broadwell a little lower.

But, if you have time, it doesn't hurt try for more, just make sure you are hitting that CPU load hard to quickly fail too high undervolt settings.

 

this is odd, every single 6700HQ i have undervolted went at least down to 150.

also you might want to repaste your system since above 90c on a 6700HQ shouldn't be possible.

 

When I get the time, maybe. But as you said, there is a point of diminishing returns and I just want a notebook that stays cool as possible but still need stability.

 

Yeah, a repaste is probably in order for this machine. What thermal paste do you recommend, the stock is IC Diamond.

 

Well, it depends on the situation, but yeah on average too high of an undervolt will fail quickly like you saw those thread failures in P95.

Running P95 long time periods in ancient lore was mostly to find power supply failures. Long run times would show PSU's that would "fold" briefly occasionally which was a problem way back when with "random" failures. I haven't seen that needed for a long time.

Like I said, the longer term hours or days later surprise BSOD's are usually from idle - non-100% load - situations when striving for and crossing one's fingers on "too high" undervolt settings.

Really it's not a competition, each CPU is unique in it's tuning, with a range of working undervolts for CPU generations known it makes it easier to zero in on the right undervolt, but there's no prize for getting the best undervolt - except maybe a Big Blue BSOD.

Again, there's not a lot of performance improvement left after -100mV vicinity is reached, you've gotten the meat of the 10c temperature drop, so you aren't thermal throttling if you ever were in the first place, so lock it in and run on it.

There is more room between -100.6 and -125mV failure point, so you can get more later if you like, when you are back into the tuning mood.

Like I said before, over time small increases in undervolt can be coaxed out of some CPU's, at -1mV or less at a time.

 

ICDiamond is fine, you can also use GCGelid or MX4.

 

If under normal use (P95 is not normal use) you aren't thermal throttling there's really no benefit to re-pasting, and it might make things worse - at the very least you open yourself up to a commitment to continual re-pastings over time as the enthusiast paste pumps or dry's out.

The whole point of tuning the software is to avoid hacking the hardware. You shouldn't need to hack the hardware if you get the software tuning locked in.

You can still adjust the fan curve, undervolt the GPU (MSI Afterburner), use RTSS to limit frame rate to the display refresh rate to stop wasting rendering time on frames dropped on the floor.

You still have more tuning available to reduce load and temperature before opening up the laptop and mucking around inside.

 

@hmscott On another one of my Clevos (same chassis, HK chip) I got a huge -144 mV UV but I just tested that using TSBench for 1 minute so I don't trust the UV, but a 65 deg C notebook while gaming with a TS profile called Melting that disabled Turbo Clocks at 65 deg. C

 

I've opened these things up many times to maintain its condition on the inside (Cleaning heatsink dust and fan dust) so I'm not really scared of mucking around

 

This thing and my other smaller Clevo chassis throttle like hell so...

 

bs advice, when a 6700HQ is running at 90+C that means the thermal paste is dry, there is absolutely no way around it, has to be done. It's similair to when you're asking yourself if you want to replace the oil in your car, there is no should or when, but refill and now.

remember then CPU runs at around 25-30W

 

Every CPU is different, you can have outliers (extremely non-average results) that will be the exception, but generally there is an established / reported range for undervolt that has been increasing in each generation.

It might be that CPU allows for a much larger stable undervolt - or it might be too aggressive and needs to be reduced to fit into the normal range of undervolt reported for that generation and model of CPU.

 

I could do the same on every laptop I own, but instead I've tuned them all with software, and returned the ones that don't tune well enough - that need a re-paste or new cooling components - and return them for another unit.

I've got better things to do with my time overall, but also I have found that I have *much* better things to do than wasting my free time pulling apart a brand new laptop and putting it back together - multiple times to get the "best" thermals.

Eventually realizing I am pulling apart all of the laptops on a regular basis, and I've now got a production line going for no good reason... no good benefit, and no rewards of tangible use.

Again, in the spirit of diminishing returns, "If it ain't broke, don't fix it" is a good motto to call into play more often than not.

 

@hmscott These things have been throttling since the day I got them years ago. It is only now I have decided to mess with finer tuning and the likes

 

Clevo's are notorious for requiring re-pasting, grinding down heating components to lay flat, flaky failing fan's, and a generally loud sucking sound from your free time disappearing into the void.

That's why I don't get Clevo's.

But, just like Alienware Tri-pod heatplates, eventually, sometimes, occasionally you get a good one out of the box, responds to software tuning, and runs fine without re-pasting.

It's worth the time to do all the tuning to see if you got a nice one.

 

That's too bad, keep tuning and see how much better they get. Maybe they'll settle in under the thermal throttling point with your fresh tuning efforts?

 

Sorry man, you really don't get it. If you do all the software tuning first and temps are good during normal use you don't *need* to re-paste.

If you re-paste will it make a difference? Does it really matter if there is no performance or functional improvement? Does re-pasting scratch the OCD-itch? Overall this is probably the main reason to do unnecessary re-pastings.

If he hasn't undervolted before he is going to have 10c lower temperatures than when he wasn't undervolted, that's a big improvement, and along with other software tunings reducing load and improving cooling it might be enough to avoid re-pasting.

If he is gauging max temperature using P95, that's not a reasonable test to use. It's a great test to test undervolt settings, with AVX / FMA disabled it's a good test to check core temperature differential, but it's a bad test to gauge day to day running temperatures, because no app is going to hit those consistent loads constantly in day to day use - besides rendering.

He can still tune the fans, undervolt the GPU - shared heatsink in this model? - use RTSS to reduce CPU/ GPU load rendering frames not displayed, which all can reduce overall temps as well.

Will he re-paste? Likely, given all the peer pressure. Does he *need* to? Probably not.

 

Well no, you don't really get it.
Even under P95, the CPU only uses around 25-30Watts. Any notebook, regardless of how bad the cooling is, should easily be able to handle that CPU even under P95 stock voltage. 98c on a 6700HQ means thermal paste is dry. If the heat doesn't transfer from CPU to heatsink, any kind of undervolt and fineruning is absolutely worthless.

 

I am not going to argue with you, as I know from previous experience you like to argue and will endlessly - I've tested it to 10 pages of round and round the same posies.

Spoiler

If you look at it from the point of view that you will do everything to tune the software to get the results you want first, you will have the best frame of mind to discover all of the software solutions to your problem that don't require hardware hacking.

If after all of the benefits of software tuning are done and you now know for a fact that you *need* hardware hacking - that there is something drastically enough wrong with your hardware that needs physical intervention - you stand on the firm ground of clearly being in control of the situation.

Otherwise you are thoughtlessly running around with screwdriver in one hand and thermal paste in the other, looking desperately for something to screw with.

I have said what I wanted to say.

Have a nice day.

 

@hmscott I don't think undervolting Maxwell Mobile GPUs is a thing...

Sent from my SM-G925F using Tapatalk

 

Hmm, I thought @Prema had a nice sideline doing custom vbios's...you'll want to consult him or check out his website:

Here's a start:
https://www.techinferno.com/index.p...ds-900m-series-overclocking-versions/&page=39

 

Well, with Unlocked vBIOS it may be possible...

Sent from my SM-G925F using Tapatalk

 

Hmm, I thought @Prema had a nice sideline doing custom vbios's...you'll want to consult him or check out his website:

Here's a start:
https://www.techinferno.com/index.p...ds-900m-series-overclocking-versions/&page=23

 

throw me a private message I can undervolt your maxwell card.

 

Disable AVX and FMA3 if you are not on a desktop.

CPUSupportsAVX=0 added in in local.txt disables BOTH AVX and FMA3
CPUSupportsFMA3=0 added in local.txt disables FMA3 but does not disable AVX.

To test the cooling solution performance only (not recommended): Small FFT.

To test VCORE stability only: Fixed FFT size (custom) of 1344K

To test RAM without destroying your BGA chips:
Custom FFT: range 512K-4096K, RAM to test: 75% of your available memory: for 32 GB this will be 24576K.

 

Is there any point to disabling FMA3 and AVX2 when AVX is already disabled? I haven't noticed any difference, that's why I only disable AVX.

 

None at all.
I just sometimes like to test with AVX on purpose, when FMA3 is just completely absurd. Windows itself and some games do use AVX instructions.
But if you're really going to test with AVX, best to do 1344K Fixed FFT's and not small FFT's.

 

I agree. Small FFTs with AVX is just too much.

 

@Falkentyne I've set min and max FFT size to 1344k in P95's custom settings. Is that how you do it?

 

Yes. That's for vcore stability testing.

 

I officially declare my 6700HQ stable, don't know about my 6820HK but my 6700HQ is stable! (During stress testing, antivirus, media consumption, and a few tabs on the web browser all at the same time)

Undervolt Stability Results - i7 6700HQ (-100.6)
-TSBench
- 64M
- 8 Threads
- Normal Priority
PASSED! (14.230 s)

-TSBench
- 256m
- 8 Threads
- Normal Priority
PASSED! (58.311 s)

-Prime95
- Small FFT 30 minutes - PASS, 0 ERRORS, 0 WARNINGS
- 1344K FFT 30 minutes - PASS, 0 ERRORS, 0 WARNINGS
- Blend 30 minutes - PASS, 0 ERRORS, 0 WARNINGS

-OCCT
- Large Data Set
- 6 min duration PASS (No BSODs so I call that a pass)
- 64 bit

@hmscott How long should I be Idle testing for?