@@yooanto9465 I wouldn't do it with a 12900k. And I run overkill cooling. The danger is probably more from long term running above temp spec than from sudden short spikes. U still have the power monitoring on the board which should perform an emergency shutdown if ur sucking down a really excessive amount of power. And I would think burning out a chip would be more dependent on the current ur feeding it than the temp it's hitting. Someone correct me if I'm wrong about that pls lol.
@@christopherjackson2157 in auto setting temp. touch 100c with arctic freezer 420II. today I disable hyperthread and C drop by 10 . Then I put all P cores to 49 and E to 39 . Voltage to 1.3 it runs way cooler now 77 c and get higher points in C23. Is this called good setting?
@@yooanto9465 that sounds much better than what you had before for sure. I don't have enough experience on alder lake to comment on specific performance numbers. 77 degrees under load is fine tho.
according to hwinfo the effective clock is "calculated across the current polling interval with respect to sleeping states" sounds to me like it is more based on usage than actual clock speed measurements. which seems to explain why it doesn't detect clock stretching
More like reported clock times logical cpu utilization (busy time) becaus it shows it "per thread" and there is no way that threads on same core could run at different clock speeds (or be in different sleep states)
In the early days of 12th gen, every processor I tried to tune and make more efficient would downclock in response to negative voltage offsets or even setting a lower fixed Vcore. I ended up giving up and running stock voltage and being limited to +100MHz core clock offsets to stay within what a 280mm AIO could cool. Thankfully, newer bios updates seem to have fixed this issue on the override and fixed vcore settings. I would update your HWINFO64; when I've encountered this issue my effective clock was expectedly lower.
Has to do with CEP I believe. Don't know all the details off the top of my head, but it's much easier to trigger when adjusting the v/f curve or having way off AC/DC ratios.
Completely agree about the "problem" with clock stretching. Spent a good hour setting up a y-cruncher stable OC just to find out that my CB score went to about stock values.
If I haven't run across this video, I swear I would've thought I was crazy! This is also very apparent with the 13th gen CPUs. I own a 13600k paired with B760i from Aorus and attempted to undervolt as the stock temps reaches 100C instantly with my 120mm AIO. I literally just threw a quick -0.1 undervolt and called it a day since HWInfo was reporting 5.1Ghz on all p-cores. Little did I know that while the core clocks are the same, the performance tanked 1/3 of what it's supposed to be! Thanks a lot for this video, as this is the first time I've heard about clock stretching as this doesn't seem to be that much of an issue with the older gen intel.
I was surprised when facing this on my 13700K too. I would say it is relatively easy to recreate this clock stretching behavior when undervolting (be it offset, A/DC load line, or LLC). I also heard that this usually happens on the AMD system when undervolting. Anyway, when clock stretching happens, the power config (e.g., voltage) is not sufficient so there is a hard limit of CPU Package Power at that setting. At 3:03 in the video (normal setting), the CPU Package Power is 210W and R23 scores as expected. Meanwhile, at 9:30, the CPU package power is only 160W when clock stretching happens. At 19:43, it is 176W. I think it is the WAIT signal that reduces power usage. From my testing, if you run at a normal power setting, but have PL1/2 capped at 160W, R23 should produce the same result. Meanwhile, if you are using a “160W clock stretching” config, but with 140W PL1/2 (lower than the max consumption). Then the system behaves as expected again, meaning the score is the same as using a normal power setting but with a 140W limitation. The closer CPU Package Power to the PL, the less clock stretching happens. That’s why whenever I undervolt the 13700k, I will make sure it can supply enough power (wattage) at every power limit (+- 10W). By the way, below is an explanation for clock stretching if anyone is interested in it. When talking about a processing unit, tasks are done by one or many instructions carried out during a clock cycle (Hz). At a very high overview level, the time of a process dictates clock speed (1/T of c). One GHz is 1 billion Hz, or 1 billion clock cycles. Clock stretching means the time to complete a cycle does not change, so your clock speed stays as is. However, the “WAIT” signal is mixed in between and thus an effective clock speed is much lower than usual. For example, with 3 Hz, you can perform the following tasks at each Hz: Start the computer, log in to OS, and open the browser. With clock stretching, the sequence becomes: Open the computer, “WAIT” (don’t do anything), login to OS. Therefore, you could not open the browser. At the hardware level, the processing unit frequency is the same as reported by sensors, but with the “WAIT” signal from firmware, it is effectively 33% slower than reported. With the same analogy (but not really about clock stretching), we have a 3600 “MHz” dual data rate (DDR) RAM stick, which is in fact 3600 MT/s. Meanwhile, the actual frequency is only 1800 MHz. DDR makes 1800 effectively become 3600 using a multiplication of 2. There are reasons for clock stretching. It will be different depending on contexts (e.g., software vs hardware), but here, it is for system stability. From an engineer's perspective, any form of protection (e.g., throttling) is better than straight-up hard crashing the system. When my 13700k clock was overstretched during an undervolting trial, even though I can run the Cinebench R23 loop for the whole day, and possibly 24/7, I lost almost 10k points. Linpack was also stable up until 200+ loops but took much longer time per loop. Prime95 errored out in the first 2 seconds for 2-3 workers, while the rest stayed strong. The same applied to my RX6800 undervolting attempt. In other words, my system actually became “fairly stable” under unstable settings.
Man that's was a long read and I'm not gonna even pretend I understood the majority of that but good to have some info regarding this. Anyway, what undervolt did you come up with? I just undervolted my 13700K by -0.1V to lower the temps a bit. I was getting similar/better Cinebench R23 results to stock (30500-31000). Weird thing was I also tried running -0.05V and for some reason I was getting lower scores (28500-29800). Wasn't clock stretching supposed to happen any voltage down from here (-0.05 or less)? Then why was my -0.1V undervolt doing better, even better than stock actually?
so what is the solution?? I have 13600k and what offset / A/DC load line or LLC settings should I use so that it wont reduce the performance by reducing the vcore only a bit?? It drivers at 1.45V at any simple game.
I found that the clock stretching usually happens when you're not stable in general. Whether due to load temperature or the load Vcore being slightly too low (not low enough for an instant crash). I guess it almost creates phantom stability because the chip might be running just fine at a reduced effective clock speed. IMO if you properly stress test and heat is accounted for, it never usually happens. It's kinda hard to achieve it. It might have something to do with setting a really conservative LLC with a much higher set Vcore. I find that if you use an LLC of 3 (auto) or 4/5 (manual OC) it usually crashes out under a Linpack or y-cruncher load way before you clock stretch. I've had clock stretching on a way too aggressive R23 run at 5.3P because the voltage was realistically way too low to run 5.3P and the temps were very bleh.
The first thing that came to mind was pulling up CPU-Z or The Intel XTU utility to check clock versus voltage reporting. Intel XTU has features I use that CPU-Z doesn't have so I tend to use it as my own verification tool. Maybe not obvious but I can't afford super expensive software that some one may have made for Professionals ...but I gotta use what I gotta use. LOL. This is weird behavior I have never seen before in my 3 decades of PC building and tinkering. This old dog learned something new. I appreciate the content an informative and educational type videos. Whether intended or not on the educational front I have learned a few things that have helped me keep a stable, voltage sipping (relative to a 9900K) overclock. I also love to keep up with the Memory Overclock and Tuning even though I do much less of that (Mostly just try and tighten timings). Cheers! Stay Healthy and Stay Sane!
About GTX 1070s "extreme undervolting" issue you mentioned @16:20 : I think it's because VRAM clock speed dropped to "2D" or "Low 3D" level, from standard full 3D clock. For whatever reason, memory clock speed just drops by 80-90% below certain vGPU value or V/F point. I have this on my big Pascal GPU as well (when tested how low I can go on vGPU).
@@ActuallyHardcoreOverclocking About CPU stretch : Can this be some weird cTDP setting at play or a power feature thing ? My logic goes something like this : MB : I'm decreasing Vcore from default by -xyz mV. CPU : Oh, you want lower vcore FROM stock... OK, sure. No problem. But once you at it, better drop my TDP to 65W/35W along with it (since we reached this [xyz] voltage drop). MB : Will do. CPU : I will show stock frequency under load, because we are on [Auto]. MB : OK.
I encountered exactly the same issue a couple of months ago, trying to push an OC by changing LLC. I used an MSI Z690, Adaptive - 0.07 offset and tested through all the LLCs. I figured out that there was a scaling of performance decrease depending on the LLC mode. Funny thing though, I could get my 12900k to around 5.5 all core with super low voltage but R23 resulted in 16k. I played around with the offset value too and found it to be not causing the issue. Only changing LLC back to auto fixed the problem.
Also happens with most mobos with offset -0.20 on locked CPUs. While up yo -0.15 you get the expected minor improvement in temperatures and scores by using less power
I noticed that this does also happen on my 12700K when undervolting with offset voltage in the BIOS Doing undervolting with Intel XTU doesn't trigger this behavior
From my limited experience with offsets on adler lake, clock stretching seems to happen when you go under VID voltage with offsets that are under max turbo ratio. Basically if go under VF point 6 voltage - you get this.
Also happens with most mobos with offset -0.20 on locked CPUs. While up to -0.15 you get the expected minor improvement in temperatures and scores by using less power
Can confirm this also happens on Ryzen, at least I can get it to happen to my 5950x. I just thought it was because it actually wasn't stable and some built in error correction was happening somewhere, my memory doesn't have ECC though..
You need to disable CEP in the BIOS for Core and SA. But that motherboard may not have access to that. The same issue is on some MSI boards. the Maximus boards have the IA CEP settings available and they are disabled by default (Auto defaults to disabled). Basically, it's based on the VID table at LOAD. Or interpolated VID. CEP wants the "die sense" vcore (Not socket sense), at full load (idk what amps rating however) to be NO LOWER than the VID at the V/F point! If it's higher it's fine. IF it's lower, it thinks your CPU is being undervolted and it causes phantom throttling (just like on X299 etc).
Reminds me of the vrm throttling on x299 which wouldn't show up in most tools as well. I don't remember but I think throttlestop was one of the few softwares picking this up.
Throttlestop also has an adaptative undervolt capability, can show if the CPU is power throttling, thermal, etc. The best tool for Intel CPUs ever made. I boosted the single core perfomance of my dinosaur i7-4700MQ almost by 10% by partially overclocking it with an adaptative undervolt through throttlestop. I gained like 200-300MHz, but i felt that difference while gaming, and microstutters in a emulator are completely gone after that, basically perfectly flat frametime now just by 300MHz more.
Also if you CPU mesh ratio ( if it's still called that for AL) needs too be hard set because the system is more likely to do this if left on AUTO. Also core temp has the ability to show you each core and not just the highest core clock, right click is your friend..
if you right click in CPU-Z you can get real time core clocks. if you have the enhanced C- states set to no limit or are allowing C5, C6 and C7 I can make my Xeon W-3175X behave the same under certain conditions. Also Core temp in the frequency section while benching will show a much lower clock. hwinfo is not good at showing the frequency in real time as it is a bit delayed.
My 5600x does this too. For example, just by setting the LLC to the lowest setting and leaving everything else on stock settings, the score is much lower. I can read that the effective clock changes in HwInfo tho. The non effective clock remains high.
hey buildzoid i love your vids even though i dont even understand most of it. i still have to ask you what you think about the upcoming msi b660 board with non k overclocking features.
Perhaps the thingy from Thermal Grizzly would help? If you're not using it already of course. IDK. 12th Gen is a mystery to me, I just hope that on Zen 4 they won't repeat the problem with only 1 lever - even though the size of the IHS is much smaller and not rectangular, I hope they won't try to save 20 cents on mounting.
I know you hate PBO but do you have any videos on it anywhere? I've been trying to figure out clock stretching on my 5600x and can't find a damn thing that I find helpful.
Yep noticed this prob when lowering voltage myself ended up just ocing to 5.2P core 41e cores @1.29v I couldn't figure out why my 12900k would run really slow when undervolted mine was doing it no matter what llc was set at on msi z690 edge
Thats strange because on TUF i put the more "efficient" option on the auto, and it manages to reduce to 1.15v on cinebench on load without any need to apply an offset
Are you doing a fixed voltage CPU undervolt or adaptative one? I always use adaptative since i don't lose perfomance and i manage to overclock a CPU by a little.
For some reason i had effective clock stretching problem on 10600k. First 10-15 min after boot cpu was running full speed and then starts stretching by 100-200mhz. Throttlestop some how fixed it by running in background.
I think you might try benching with some trivial C programs designed to measure wall clock while doing something that will only ever touch L1 cache. If you have access to cygwin/wsl 1 (not 2!)/linux, you could run something like: gcc -O0 -x c - -o aout
Does this behavior change with different "SVID behavior" settings? I think you can reduce the voltage by using Trained or Best Case Scenario mode. Also, do you set AC LL and DC LL along with your LLC?
Is CEP enabled? Static offset will trip CEP and cause a reduction in performance depending on the offset, which is why you should use VF offset mode instead.
You can easily replicate this using CPU-Z stress too. BTW Happens with AMD Ryzen also. If there is digital readout on the motherboard you could check there too ?
Have you tried ram overclocking with non-K 12th gen parts? I know edge use case but I can't even enable xmp with my 12400f in the MSI Pro B660M A and some people are suggesting it's because the memory controller is supposedly voltage locked on non-K parts
Is this literally clock stretching (in the same sense that slow I2C devices hold the clock low and change the clock’s duty cycle) or is it just a change in frequency? Do CPUs have pins that output a signal that tracks the real clock frequency/phase? There’s gotta be a debugging mode for this
Have you tried a Another power plan example bitsum highest performance power plan From my experience solves many of these problems Especially in y cruncher
Hey guys. I have an idea that won't leave me. So my motherboard MSI B450-a PRO has newer version MSI B450-a PRO MAX. I belive the only difference is that it has 32mb bios chip instead of 16mb. I wonder if I can expect correct functioning of the board if I just replace chip to a new with a new bios? The only reason is better UEFI UI and built in OC profiles memory instead of using USB stick. What do you think?
Honestly I'm no where near knowledgeable to answer if that will make a difference though. However though, if you're eager enough to try, just find out the exact 32MB BIOS chip that is used in the newer version and look for it on AliExpress/eBay or anywhere else it can be sourced from. Second you'll need a BIOS programmer, personally I bought 'CH341A' from AliExpress for really cheap and it works like a charm. Then simply program the BIOS onto the cheap using software like AsProgrammer. Next, replace the BIOS chip in the motherboard with the new one. The BIOS programmer is a very effective tool to have regardless, in case your family or friends soft brick their motherboard or GPU, it should be relatively easy to burn onto the chip a new BIOS using it.
G'day Buildzoid, So just wondering if this is the effect GN Steve has spoken about where you need to test your results because there are many comments everywhere with people claiming... "Just Underclock & you use less power at the same clocks or without hurting performance"
This is pretty severe stretching. I experienced clock stretching on the order of maybe 10-20% performance loss on my ryzen 5900x using a voltage offset. This is a whole other level lol At least on Ryzen HWinfo's effective clock reading shows that you are clock stretching.
@@nickpowell7432 In hwinfo you will see values for clock speed and effective clock speed. Effective clock speed will be lower than clock speed if your cpu is stretching. But as seen in this video that isn’t always the case. You need to verify with a benchmark. However, if your effective is reporting lower than clock speed you can be 99% sure you are indeed stretching.
interesting.. i have similar problem with my 1080ti HOF. runs max with 2139mhz / 1.35v... stable and for all day 2088mhz / 1.21 v. If I stay on a certain clock and just give more or less voltage, the card will slower or faster, weird thing, never knew why
I don't know if my 12900k is clock stretching or not, because I am using aorus pro and I have set the llc to low at day one, but I have never reached over 28k points in cb23 even in I am in 4.9ghz all core and 4.0ghz e core. I am now trying to set it back to normal llc or auto to see if it can reach higher score, but I think the voltage would be much higher since I am doing 5.2 ghz all core right now :\
What's really bizarre is that, if it does this now, what's to say its not always misreporting the clock speed? How does the user know, beyond the shadow of a doubt, what the clock speed is actually is? There is no way to accurately determine what the actual clock speed is.
You can't trust HWinfo64 effective clock reporting. On my 5800x it tells me one of the cores reached 5544 MHz while the global frequency limit max is 4925 MHz.
My 12900k is ruining all core 5ghz and 4ghz on e cores, with an offset of minus 0.0140 and it runs perfectly never going above 60c under full load. Using windows 11 though.
If it takes a certain amount of voltage to reach a frequency, then that is the voltage the CPU needs. Anything above will increase the clocks, anything below will drop that clock speed. That is how fixed voltage values work and it doesn't bluescreen. That's why peoplr complained about PBO from AMD since it can't do an adaptative undervolt. When you do an adaptative undervolt (PBO 2 or Curve optimizer, Throttlestop and XTU), things are different, when you use a small number of cores, there is higher headroom to take out voltage to reach the same frequencies, therefore you gain OC headroom vs stock perfomance. Then at base clock, there's less headroom to take voltage out of the base clocks since the CPU cores are being totally utilized and require more power instead of the CPU using a small number of threads and less power. If you enable the power limits and the e cores (you let the e cores doing their job and the P cores reaching base clock when they should), you should get in theory better Cinebench R20 perfomance with an adaptative offset and better efficency since core count is better in that benchmark than a small bump in clockspeed. Ryzen 9 5950X with 16 perfomance cores at 3.8GHz easily beats an e core disabled i9-12900K at 4.9GHz.
This is so weird, my dud of an 12700KF can do 4.9 all core at 1.25V llc3 which is 1.236 on load without compromising the performance by much. I9 should be a better bin 🤔
I wondered why I was getting 500 points higher in cpu timespy scores compared to other Similar cpus but they had faster clocks reported but lower scores.
lol I currently have an i7-9700 paired with a cheapo Gigabyte B360M DS3H that used to run with an i5-8400 previously, I used ThrottleStop to undervolt the CPU, I selected CPU Core, CPU Cache, iGPU CPU > Set it to Adaptive > Unlocked voltages > Set an offset of -0.7v on all three, cinebench didn't crash, I opened and played games, no problem, so I tried putting the offset at -1.0v, cinebench didn't crash and it was going well, it ended the process but then all my apps started saying "Not responding" until PC crashed ( Blue screen ) , PC restarted > Logged in > Immediately again blue screen > Restarted > Logged in > reset values to stock real quick, it was kinda scary, guess Il stay with -0.07v for the time being, testing these things because I'm going to purchase an i7-12700K and pair it with an MSI B660M Mag Mortar DDR4 WIFI maybe next month, I live in a hot country and my room has bad airflow (dont know the ambient temp but its prob quite warm), at idle with a 240mm corsair aio I get like 40-43C idle and reaches 75C max (which is not bad) but that's a 65W TDP locked CPU, im getting an i7-12700K (No OC) but should be at least like 10C higher I guess so idk that's why Im researching about undervolting. gonna get three 120mm case fans that can ramp up at 2000rpm if needed to put as front intake just in case.
Ahh clever. I'd always figured that was just instability causing lower scores than the core speed would warrant. Good video. Specifically on bulldozer I'll hit a point where the scores basically all of a sudden stop going up and get cut to almost exactly half what they were. I assume that's something similar?
I suppose I've got a good sample then, as I get 10.5k in CB R20 at just 1.25V LLC3 fixed voltage, been running it for a few months and never had a single problem. Very stable platform, at least for me. MSI Tomahawk.
That seems to be just 100% normal behavior? Caused by instability? CPUs have had error correction strategies for a long time already. When they start getting unstable they will likely "redo" command/transfers/etc and just effectively get slower. When they get unstable enough they can't recover, and will crash.
Actually Hardcore Clock Stretching
that one particular "l" is doing a lot of work in that phrase
😂❤
24:56 - the best part of the video right here
If you disable thermal monitoring, the CPU stops throttling under load!
hell yeah
Lol yea I do that almost by default now
If it's too hot I'll figure it out when it crashes :)
What if it burns out?
@@yooanto9465 I wouldn't do it with a 12900k. And I run overkill cooling. The danger is probably more from long term running above temp spec than from sudden short spikes. U still have the power monitoring on the board which should perform an emergency shutdown if ur sucking down a really excessive amount of power. And I would think burning out a chip would be more dependent on the current ur feeding it than the temp it's hitting. Someone correct me if I'm wrong about that pls lol.
@@christopherjackson2157 in auto setting temp. touch 100c with arctic freezer 420II.
today I disable hyperthread and C drop by 10 . Then I put all P cores to 49 and E to 39 . Voltage to 1.3 it runs way cooler now 77 c and get higher points in C23. Is this called good setting?
@@yooanto9465 that sounds much better than what you had before for sure. I don't have enough experience on alder lake to comment on specific performance numbers.
77 degrees under load is fine tho.
according to hwinfo the effective clock is "calculated across the current polling interval with respect to sleeping states"
sounds to me like it is more based on usage than actual clock speed measurements. which seems to explain why it doesn't detect clock stretching
More like reported clock times logical cpu utilization (busy time) becaus it shows it "per thread" and there is no way that threads on same core could run at different clock speeds (or be in different sleep states)
how did you comment 4 days ago??
@@standardaccountenjoyer2826 unlisted
@@standardaccountenjoyer2826 Probably patreon early access too?
In the early days of 12th gen, every processor I tried to tune and make more efficient would downclock in response to negative voltage offsets or even setting a lower fixed Vcore. I ended up giving up and running stock voltage and being limited to +100MHz core clock offsets to stay within what a 280mm AIO could cool. Thankfully, newer bios updates seem to have fixed this issue on the override and fixed vcore settings. I would update your HWINFO64; when I've encountered this issue my effective clock was expectedly lower.
Has to do with CEP I believe. Don't know all the details off the top of my head, but it's much easier to trigger when adjusting the v/f curve or having way off AC/DC ratios.
Completely agree about the "problem" with clock stretching. Spent a good hour setting up a y-cruncher stable OC just to find out that my CB score went to about stock values.
If I haven't run across this video, I swear I would've thought I was crazy! This is also very apparent with the 13th gen CPUs. I own a 13600k paired with B760i from Aorus and attempted to undervolt as the stock temps reaches 100C instantly with my 120mm AIO. I literally just threw a quick -0.1 undervolt and called it a day since HWInfo was reporting 5.1Ghz on all p-cores. Little did I know that while the core clocks are the same, the performance tanked 1/3 of what it's supposed to be!
Thanks a lot for this video, as this is the first time I've heard about clock stretching as this doesn't seem to be that much of an issue with the older gen intel.
I was surprised when facing this on my 13700K too. I would say it is relatively easy to recreate this clock stretching behavior when undervolting (be it offset, A/DC load line, or LLC). I also heard that this usually happens on the AMD system when undervolting. Anyway, when clock stretching happens, the power config (e.g., voltage) is not sufficient so there is a hard limit of CPU Package Power at that setting. At 3:03 in the video (normal setting), the CPU Package Power is 210W and R23 scores as expected. Meanwhile, at 9:30, the CPU package power is only 160W when clock stretching happens. At 19:43, it is 176W. I think it is the WAIT signal that reduces power usage. From my testing, if you run at a normal power setting, but have PL1/2 capped at 160W, R23 should produce the same result. Meanwhile, if you are using a “160W clock stretching” config, but with 140W PL1/2 (lower than the max consumption). Then the system behaves as expected again, meaning the score is the same as using a normal power setting but with a 140W limitation. The closer CPU Package Power to the PL, the less clock stretching happens. That’s why whenever I undervolt the 13700k, I will make sure it can supply enough power (wattage) at every power limit (+- 10W).
By the way, below is an explanation for clock stretching if anyone is interested in it.
When talking about a processing unit, tasks are done by one or many instructions carried out during a clock cycle (Hz). At a very high overview level, the time of a process dictates clock speed (1/T of c). One GHz is 1 billion Hz, or 1 billion clock cycles. Clock stretching means the time to complete a cycle does not change, so your clock speed stays as is. However, the “WAIT” signal is mixed in between and thus an effective clock speed is much lower than usual. For example, with 3 Hz, you can perform the following tasks at each Hz: Start the computer, log in to OS, and open the browser. With clock stretching, the sequence becomes: Open the computer, “WAIT” (don’t do anything), login to OS. Therefore, you could not open the browser. At the hardware level, the processing unit frequency is the same as reported by sensors, but with the “WAIT” signal from firmware, it is effectively 33% slower than reported. With the same analogy (but not really about clock stretching), we have a 3600 “MHz” dual data rate (DDR) RAM stick, which is in fact 3600 MT/s. Meanwhile, the actual frequency is only 1800 MHz. DDR makes 1800 effectively become 3600 using a multiplication of 2.
There are reasons for clock stretching. It will be different depending on contexts (e.g., software vs hardware), but here, it is for system stability. From an engineer's perspective, any form of protection (e.g., throttling) is better than straight-up hard crashing the system. When my 13700k clock was overstretched during an undervolting trial, even though I can run the Cinebench R23 loop for the whole day, and possibly 24/7, I lost almost 10k points. Linpack was also stable up until 200+ loops but took much longer time per loop. Prime95 errored out in the first 2 seconds for 2-3 workers, while the rest stayed strong. The same applied to my RX6800 undervolting attempt. In other words, my system actually became “fairly stable” under unstable settings.
Man that's was a long read and I'm not gonna even pretend I understood the majority of that but good to have some info regarding this. Anyway, what undervolt did you come up with?
I just undervolted my 13700K by -0.1V to lower the temps a bit. I was getting similar/better Cinebench R23 results to stock (30500-31000). Weird thing was I also tried running -0.05V and for some reason I was getting lower scores (28500-29800). Wasn't clock stretching supposed to happen any voltage down from here (-0.05 or less)? Then why was my -0.1V undervolt doing better, even better than stock actually?
so what is the solution?? I have 13600k and what offset / A/DC load line or LLC settings should I use so that it wont reduce the performance by reducing the vcore only a bit?? It drivers at 1.45V at any simple game.
@@Golden2Talon Just run override mode all core 5.1ghz with little llc vdrop like 1.2v idle ->1.16v prime95avx.
You can use throttlestop to control clock stretching. Dont know if it itll prevent it but you can definitely induce it.
I found that the clock stretching usually happens when you're not stable in general. Whether due to load temperature or the load Vcore being slightly too low (not low enough for an instant crash). I guess it almost creates phantom stability because the chip might be running just fine at a reduced effective clock speed.
IMO if you properly stress test and heat is accounted for, it never usually happens. It's kinda hard to achieve it. It might have something to do with setting a really conservative LLC with a much higher set Vcore.
I find that if you use an LLC of 3 (auto) or 4/5 (manual OC) it usually crashes out under a Linpack or y-cruncher load way before you clock stretch. I've had clock stretching on a way too aggressive R23 run at 5.3P because the voltage was realistically way too low to run 5.3P and the temps were very bleh.
The first thing that came to mind was pulling up CPU-Z or The Intel XTU utility to check clock versus voltage reporting. Intel XTU has features I use that CPU-Z doesn't have so I tend to use it as my own verification tool. Maybe not obvious but I can't afford super expensive software that some one may have made for Professionals ...but I gotta use what I gotta use. LOL. This is weird behavior I have never seen before in my 3 decades of PC building and tinkering. This old dog learned something new. I appreciate the content an informative and educational type videos. Whether intended or not on the educational front I have learned a few things that have helped me keep a stable, voltage sipping (relative to a 9900K) overclock. I also love to keep up with the Memory Overclock and Tuning even though I do much less of that (Mostly just try and tighten timings). Cheers! Stay Healthy and Stay Sane!
Great vid Buildzoid, like always!
About GTX 1070s "extreme undervolting" issue you mentioned @16:20 :
I think it's because VRAM clock speed dropped to "2D" or "Low 3D" level, from standard full 3D clock. For whatever reason, memory clock speed just drops by 80-90% below certain vGPU value or V/F point. I have this on my big Pascal GPU as well (when tested how low I can go on vGPU).
Nvidia cards clock strech across pretty much the entire voltage range. The less voltage you run the more severe it gets.
@@ActuallyHardcoreOverclocking About CPU stretch : Can this be some weird cTDP setting at play or a power feature thing ?
My logic goes something like this :
MB : I'm decreasing Vcore from default by -xyz mV.
CPU : Oh, you want lower vcore FROM stock... OK, sure. No problem.
But once you at it, better drop my TDP to 65W/35W along with it (since we reached this [xyz] voltage drop).
MB : Will do.
CPU : I will show stock frequency under load, because we are on [Auto].
MB : OK.
I encountered exactly the same issue a couple of months ago, trying to push an OC by changing LLC. I used an MSI Z690, Adaptive - 0.07 offset and tested through all the LLCs. I figured out that there was a scaling of performance decrease depending on the LLC mode. Funny thing though, I could get my 12900k to around 5.5 all core with super low voltage but R23 resulted in 16k.
I played around with the offset value too and found it to be not causing the issue. Only changing LLC back to auto fixed the problem.
Also happens with most mobos with offset -0.20 on locked CPUs. While up yo -0.15 you get the expected minor improvement in temperatures and scores by using less power
I noticed that this does also happen on my 12700K when undervolting with offset voltage in the BIOS
Doing undervolting with Intel XTU doesn't trigger this behavior
From my limited experience with offsets on adler lake, clock stretching seems to happen when you go under VID voltage with offsets that are under max turbo ratio. Basically if go under VF point 6 voltage - you get this.
Also happens with most mobos with offset -0.20 on locked CPUs. While up to -0.15 you get the expected minor improvement in temperatures and scores by using less power
i tried on i5-12500 and on -0.05 perf decreased by 50%. on -0.1 it decreased by about 68%
it seems to me that it's behaving like x299 "phantom throttling"
actually yeah it is very similar to that.
Love 💘 this video 📹 BZ! Keep up the great work 👍!
Can confirm this also happens on Ryzen, at least I can get it to happen to my 5950x. I just thought it was because it actually wasn't stable and some built in error correction was happening somewhere, my memory doesn't have ECC though..
yes zen cpus also clock stretch
Btw why is the video tagged with "z690" but its really a B660-F ?
You need to disable CEP in the BIOS for Core and SA.
But that motherboard may not have access to that.
The same issue is on some MSI boards.
the Maximus boards have the IA CEP settings available and they are disabled by default (Auto defaults to disabled).
Basically, it's based on the VID table at LOAD. Or interpolated VID. CEP wants the "die sense" vcore (Not socket sense), at full load (idk what amps rating however) to be NO LOWER than the VID at the V/F point! If it's higher it's fine. IF it's lower, it thinks your CPU is being undervolted and it causes phantom throttling (just like on X299 etc).
Can you specify what CEP is? how should one found it in BIOS, thanks?
Reminds me of the vrm throttling on x299 which wouldn't show up in most tools as well. I don't remember but I think throttlestop was one of the few softwares picking this up.
Throttlestop is designed for laptop users dealing with terrible motherboards and cooling since many, many years ago 😂
Throttlestop also has an adaptative undervolt capability, can show if the CPU is power throttling, thermal, etc. The best tool for Intel CPUs ever made.
I boosted the single core perfomance of my dinosaur i7-4700MQ almost by 10% by partially overclocking it with an adaptative undervolt through throttlestop. I gained like 200-300MHz, but i felt that difference while gaming, and microstutters in a emulator are completely gone after that, basically perfectly flat frametime now just by 300MHz more.
Seems this is a Asus Problem not an 12th/13th gen Problem . On Msi u have Cpu Liteload Control.
Also if you CPU mesh ratio ( if it's still called that for AL) needs too be hard set because the system is more likely to do this if left on AUTO. Also core temp has the ability to show you each core and not just the highest core clock, right click is your friend..
Hardcore clock and core torture. Please put my hardware in the clock stretching machine.
if you right click in CPU-Z you can get real time core clocks. if you have the enhanced C- states set to no limit or are allowing C5, C6 and C7 I can make my Xeon W-3175X behave the same under certain conditions. Also Core temp in the frequency section while benching will show a much lower clock. hwinfo is not good at showing the frequency in real time as it is a bit delayed.
My 5600x does this too. For example, just by setting the LLC to the lowest setting and leaving everything else on stock settings, the score is much lower. I can read that the effective clock changes in HwInfo tho. The non effective clock remains high.
interesting, thanks.
What is clock stretching?
hey buildzoid i love your vids even though i dont even understand most of it. i still have to ask you what you think about the upcoming msi b660 board with non k overclocking features.
My 12400F clockstretches on Asus B660m K if I lower voltage offset by more than -15mv in bios.
Exactly the same here and also with gigabyte boards
Perhaps the thingy from Thermal Grizzly would help? If you're not using it already of course. IDK. 12th Gen is a mystery to me, I just hope that on Zen 4 they won't repeat the problem with only 1 lever - even though the size of the IHS is much smaller and not rectangular, I hope they won't try to save 20 cents on mounting.
"you couldn't Bluescreen" Yes you could, but you'd only notice the next day. LOL
Actually Hardcore Underclocking
I know you hate PBO but do you have any videos on it anywhere? I've been trying to figure out clock stretching on my 5600x and can't find a damn thing that I find helpful.
I presume I'm missing something and manually setting static VCore doesn't stop this behaviour?
Yep noticed this prob when lowering voltage myself ended up just ocing to 5.2P core 41e cores @1.29v I couldn't figure out why my 12900k would run really slow when undervolted mine was doing it no matter what llc was set at on msi z690 edge
Thats strange because on TUF i put the more "efficient" option on the auto, and it manages to reduce to 1.15v on cinebench on load without any need to apply an offset
Are you doing a fixed voltage CPU undervolt or adaptative one?
I always use adaptative since i don't lose perfomance and i manage to overclock a
CPU by a little.
For some reason i had effective clock stretching problem on 10600k. First 10-15 min after boot cpu was running full speed and then starts stretching by 100-200mhz. Throttlestop some how fixed it by running in background.
I think you might try benching with some trivial C programs designed to measure wall clock while doing something that will only ever touch L1 cache. If you have access to cygwin/wsl 1 (not 2!)/linux, you could run something like:
gcc -O0 -x c - -o aout
Does the Max Core Frequency and the Processor Cache Frequency display correctly in the Intel Extreme Tuning Utility during clock stretching?
Does this behavior change with different "SVID behavior" settings? I think you can reduce the voltage by using Trained or Best Case Scenario mode. Also, do you set AC LL and DC LL along with your LLC?
Is CEP enabled? Static offset will trip CEP and cause a reduction in performance depending on the offset, which is why you should use VF offset mode instead.
Ok, that's fun.
Silicone is showing the game is nearly up!
This also show up on cpuz testing.
So that's probably why my linpack score has been terrible and i had no idea what was wrong.
You can easily replicate this using CPU-Z stress too. BTW Happens with AMD Ryzen also.
If there is digital readout on the motherboard you could check there too ?
I enthusiastically clicked this video, but then I realised I read the title wrong, clock stretching. I tought Intel started a new line of products.
Have you tried ram overclocking with non-K 12th gen parts? I know edge use case but I can't even enable xmp with my 12400f in the MSI Pro B660M A and some people are suggesting it's because the memory controller is supposedly voltage locked on non-K parts
yes VCCSA is locked and the CPUs max out between 3200 and 3600 in gear1 and 4533 to 5200 in gear2
Is this literally clock stretching (in the same sense that slow I2C devices hold the clock low and change the clock’s duty cycle) or is it just a change in frequency? Do CPUs have pins that output a signal that tracks the real clock frequency/phase? There’s gotta be a debugging mode for this
Have you tried disabling CEP to see if the problem persist?
Have you tried a Another power plan
example
bitsum highest performance power plan
From my experience solves many of these problems
Especially in y cruncher
Well, my 12100f works at -0.2V offset without a problem. At -0.23V it just crashes under load. Maybe this happens only with higher clocked CPUs.
Hey guys. I have an idea that won't leave me.
So my motherboard MSI B450-a PRO has newer version MSI B450-a PRO MAX. I belive the only difference is that it has 32mb bios chip instead of 16mb. I wonder if I can expect correct functioning of the board if I just replace chip to a new with a new bios? The only reason is better UEFI UI and built in OC profiles memory instead of using USB stick.
What do you think?
Honestly I'm no where near knowledgeable to answer if that will make a difference though. However though, if you're eager enough to try, just find out the exact 32MB BIOS chip that is used in the newer version and look for it on AliExpress/eBay or anywhere else it can be sourced from. Second you'll need a BIOS programmer, personally I bought 'CH341A' from AliExpress for really cheap and it works like a charm. Then simply program the BIOS onto the cheap using software like AsProgrammer. Next, replace the BIOS chip in the motherboard with the new one. The BIOS programmer is a very effective tool to have regardless, in case your family or friends soft brick their motherboard or GPU, it should be relatively easy to burn onto the chip a new BIOS using it.
hi BZ how many DEL\F2 keys do you have replace a year? 🤭
G'day Buildzoid,
So just wondering if this is the effect GN Steve has spoken about where you need to test your results because there are many comments everywhere with people claiming...
"Just Underclock & you use less power at the same clocks or without hurting performance"
This is pretty severe stretching. I experienced clock stretching on the order of maybe 10-20% performance loss on my ryzen 5900x using a voltage offset. This is a whole other level lol
At least on Ryzen HWinfo's effective clock reading shows that you are clock stretching.
What do you look for to see if it’s stretching?
@@nickpowell7432 In hwinfo you will see values for clock speed and effective clock speed. Effective clock speed will be lower than clock speed if your cpu is stretching. But as seen in this video that isn’t always the case. You need to verify with a benchmark. However, if your effective is reporting lower than clock speed you can be 99% sure you are indeed stretching.
@@rdiznfriends I thought that’s what you were going to say but I wasn’t sure thanks.
do you run on dual rank mode or single rank
interesting.. i have similar problem with my 1080ti HOF. runs max with 2139mhz / 1.35v... stable and for all day 2088mhz / 1.21 v. If I stay on a certain clock and just give more or less voltage, the card will slower or faster, weird thing, never knew why
I don't know if my 12900k is clock stretching or not, because I am using aorus pro and I have set the llc to low at day one, but I have never reached over 28k points in cb23 even in I am in 4.9ghz all core and 4.0ghz e core. I am now trying to set it back to normal llc or auto to see if it can reach higher score, but I think the voltage would be much higher since I am doing 5.2 ghz all core right now :\
What about 12600kf?
Any particular reason you're using R20 over R23?
doesn't R23 take longer?
R23 does take longer.
I think R20 is more CPU intensive than R23 because the optimization is worse and the processors get hotter.
Does this behaviour appear on Windows 11.
Zen+, Zen2&3 do have clock streching too....I'm not sure about Zen.
It's a very strange "Feature"!
Is that a bug in the CPU microcode?
What's really bizarre is that, if it does this now, what's to say its not always misreporting the clock speed? How does the user know, beyond the shadow of a doubt, what the clock speed is actually is? There is no way to accurately determine what the actual clock speed is.
holy shit 64GB RAM, 4x16 bz?
You can do the same with Ryzen. I was trying undervolt and got same results.
You can't trust HWinfo64 effective clock reporting. On my 5800x it tells me one of the cores reached 5544 MHz while the global frequency limit max is 4925 MHz.
Does this ruin the new Hwbot competition - can just run 6ghz validated, while in actuallty running 3Ghz?
Reminds me of my 5800X...everything above 4.85 GHz = clock stretching.
My 12900k is ruining all core 5ghz and 4ghz on e cores, with an offset of minus 0.0140 and it runs perfectly never going above 60c under full load. Using windows 11 though.
I thought it's the year 2022 your computer is so advanced it's jumped a year😆😆😆😆😆
Is this a 12th gen issue or is it unique to ASUS ?
You have the e-cores disabled.
i think im getting this on my 14900k if i dont give it some xtra voltage
I have the same problem with my 13700K.
If it takes a certain amount of voltage to reach a frequency, then that is the voltage the CPU needs.
Anything above will increase the clocks, anything below will drop that clock speed.
That is how fixed voltage values work and it doesn't bluescreen. That's why peoplr complained about PBO from AMD since it can't do an adaptative undervolt.
When you do an adaptative undervolt (PBO 2 or Curve optimizer, Throttlestop and XTU), things are different, when you use a small number of cores, there is higher headroom to take out voltage to reach the same frequencies, therefore you gain OC headroom vs stock perfomance.
Then at base clock, there's less headroom to take voltage out of the base clocks since the CPU cores are being totally utilized and require more power instead of the CPU using a small number of threads and less power.
If you enable the power limits and the e cores (you let the e cores doing their job and the P cores reaching base clock when they should), you should get in theory better Cinebench R20 perfomance with an adaptative offset and better efficency since core count is better in that benchmark than a small bump in clockspeed. Ryzen 9 5950X with 16 perfomance cores at 3.8GHz easily beats an e core disabled i9-12900K at 4.9GHz.
This is so weird, my dud of an 12700KF can do 4.9 all core at 1.25V llc3 which is 1.236 on load without compromising the performance by much. I9 should be a better bin 🤔
also happens to 9th gen !
Nothing like that up to 12 gen. And i undervolt pretty much every computer that comes through my lab.. a couple every day
I wondered why I was getting 500 points higher in cpu timespy scores compared to other Similar cpus but they had faster clocks reported but lower scores.
the timespy CPU test is incredibly sensitive to RAM settings.
lol I currently have an i7-9700 paired with a cheapo Gigabyte B360M DS3H that used to run with an i5-8400 previously, I used ThrottleStop to undervolt the CPU, I selected CPU Core, CPU Cache, iGPU CPU > Set it to Adaptive > Unlocked voltages > Set an offset of -0.7v on all three, cinebench didn't crash, I opened and played games, no problem, so I tried putting the offset at -1.0v, cinebench didn't crash and it was going well, it ended the process but then all my apps started saying "Not responding" until PC crashed ( Blue screen ) , PC restarted > Logged in > Immediately again blue screen > Restarted > Logged in > reset values to stock real quick, it was kinda scary, guess Il stay with -0.07v for the time being, testing these things because I'm going to purchase an i7-12700K and pair it with an MSI B660M Mag Mortar DDR4 WIFI maybe next month, I live in a hot country and my room has bad airflow (dont know the ambient temp but its prob quite warm), at idle with a 240mm corsair aio I get like 40-43C idle and reaches 75C max (which is not bad) but that's a 65W TDP locked CPU, im getting an i7-12700K (No OC) but should be at least like 10C higher I guess so idk that's why Im researching about undervolting. gonna get three 120mm case fans that can ramp up at 2000rpm if needed to put as front intake just in case.
intel hot only i7 8700k perfect cpu, O watercooling 1.33v 85C intel sack
Your on W10 you need to set CPU priority to above normal so even when not in focus you get full performance.
You're*
Maybe it has on-die ECC 😁
Ahh clever. I'd always figured that was just instability causing lower scores than the core speed would warrant. Good video.
Specifically on bulldozer I'll hit a point where the scores basically all of a sudden stop going up and get cut to almost exactly half what they were. I assume that's something similar?
I suppose I've got a good sample then, as I get 10.5k in CB R20 at just 1.25V LLC3 fixed voltage, been running it for a few months and never had a single problem. Very stable platform, at least for me. MSI Tomahawk.
So if you undervolt this is now the behaviour of modern hardware?? o.o
Only with Ryzen and 12 gen Intel. Also Ryzen has curve optimizer that does real undervolt
Not crashing AND producing correct results is always preferable.
That seems to be just 100% normal behavior? Caused by instability? CPUs have had error correction strategies for a long time already. When they start getting unstable they will likely "redo" command/transfers/etc and just effectively get slower. When they get unstable enough they can't recover, and will crash.
Not first :^)