The ThrottleStop Guide

Thanks for your response. My only aim was to lower the temps as much as I can. Thats why I wanted to try -300 mV.
Generally I limit my cpu to 3.2ghz for gaming since my gtx1060 is the bottleneck factor. I dont like the fan sound above 80 degrees thus I rarely use 3.9ghz. Applying Undervolt and 3.2ghz limit my temps hover around 70-75 degrees for both Cpu and Gpu in AAA games with a better fan noise.
Btw i plan to give my laptop to service for cleaning and repasting(this model sucks to open so i don't bother myself). While in service they can also update my bios to latest available. Do you know if tongfang also patches the undervolting like dell and msi did?

 

I would give my first born away before I would give my laptop to a service center to do a BIOS update. Is there something wrong with your computer that it needs a BIOS update? In 2020, this procedure is definitely not worth the risk.

Your laptop is running great and you have complete control over how it runs. Why would you want to chance a BIOS update? You should also learn to clean and re-paste your own laptop. It is not rocket science. No one cares more about your laptop than you.

 

I wouldn't bother with bios update than. Not work the risk indeed.

As for cleaning I did repaste my old laptop which was just remove bottom cover and be done with it. This one however open from top to bottom with a lot of cables to disconnect. Only disassembly video I found takes 1 hour(ofc he is just a regular user but still its tedious). Here's the link

 

Did some research on the cache/core voltage link/sync after seeing some folks actually believing that they are running a negative core voltage offset of up to 1000mV on KBL-R and later chips (where Intel began syncing cache and core voltage domains by default).

I found that setting a lower Core vs. Cache voltage did in fact not change the core voltage domain at all. It always used the cache voltage as the actual limit for both domains.

But, what it actually did was reducing the automatically (by most microcodes) and additionally applied overvoltage during AVX2 instruction set workloads.

While this exact voltage depends on each ASIC, the average automatic AVX2 overvoltage is around +100mV on top of the regular instruction set voltage.
By setting up to a negative -1000mV (TS max) on core, the effective AVX2 overvoltage has been reduced by up to -75mV effectively bringing it down to only +25mV above the regular workload voltage.

Makes sense as AVX2 ratio clipping was introduced alongside the 8th gen that began linking the voltage domains.

Intel must have either planned to be able to regulate the AVX2 overvoltage offset separately from regular Cache and Core voltage and then left it hidden under the hood or we are loopholing into the AVX2 ratio clipping and forcing the core to drop a couple of voltage domains without the intended core clock drop.

Maybe that helps some people to understand that, no we are not undervolting the core higher than the cache (as it used to be ≤ 7th gen), we are reducing, though not eliminating, the additional AVX2 overvoltage.

Just my 2c

 

I don't know if this is a reliable test in any way, but the screenshot below shows CPU-Z AVX TEST with core and cache linked:


and another one with core at -1000mV



As you can see there is a frequency difference for a -1000mV setting from a start of the cpu-z test. Of course the clock are droping down with the rising temp but with -1000mV the clocks are higher.

So it means that every app/game and so on that uses avx instructions will benefit form it, that is way some of us were seeing drops in temps.
Although for some reason my 8th gen refresh U CPU cannot maintain full all core boost with avx instructions in this test while on non avx there is no problem to reach max all core 4.1GHz boost.

So if around -1000 in TS gave us -75mV on AVX as @Prema said my question is to @unclewebb if we can rise the -1000mV limit to fully remove avx overvoltage?

I will retest this on my 10875H when I will be at home.

Best regards

 

Setting the core offset request higher than the cache offset has been recommended for a while. The @Prema explanation above makes sense. Requesting different voltages tricks the CPU so it reduces the amount of overvoltage it uses when it encounters AVX2 instructions. I have been recommending running CInebench R20 to prove this to users. It heavily uses the AVX2 instructions so one should see better temps or better performance after setting the voltages to different values.

If CPU-Z heavily uses AVX2 instructions then this should be another convenient way to prove that using different voltage offsets for core and cache is a good thing to be doing.

I am not yet sure if the 8th, 9th and 10th Gen CPUs all use the same amount of AVX2 overvoltage (+100 mV) and if it is only possible to remove +75 mV of that. Some 8750H owners have reported their best results with the offset voltages set close to these values.

cache offset = -125 mV
core offset = -225 mV

Maybe the 8th Gen can completely remove all of the +100 mV AVX2 overvolt when set like this. Hopefully CPU-Z proves to be a useful test and users do some hands on testing with different CPUs to try and prove this.

A setting of -1000 mV is not necessary. As long as the core is set at least -100 mV more than the cache, any AVX2 overvoltage should be eliminated. Setting the core offset to -200 mV more or -500 mV or -1000 mV more than the cache makes no difference.

@Krzyslaw - Your first test shows a consistent 36 multiplier. When 4 cores are active, the 8565U can use a 41 multiplier. Try opening up Limit Reasons while testing to see if there is any reason for the throttling that is in progress. Your CPU is well below the 99°C thermal throttling temperature that your laptop is set to. Make sure HWiNFO in not running in the background when using Limit Reasons. Is Speed Shift Max set to at least 46?

 

I finally have access to a 10th Gen Core i9-10850K. This should allow me to add a few more features to TS to better support these 10 core beasts. Let the testing begin.

For all tests, I am running 50X on all cores, 46X on the cache, Speed Shift EPP = 0 for max CPU speed when idle. I will be using Cinebench R20 and the TS screenshots will be taken after Cinebench completes.

Core 0 mV, Cache 0 mV
279.0 W, 94°C


Yikes!!! These things really do put out a lot of heat. The default voltage is always a little on the high side so time to use ThrottleStop to do some undervolting. There is an ongoing debate about whether the core and cache should be adjusted equally. Proper science involves forgetting about everything that one knows or thinks they know and go into any testing blind.

Core -50 mV, Cache 0 mV
249.7 W, 86°C


After adjusting only the core offset, there is already a big drop in reported power consumption and the peak CPU package temperature confirms this.

Core -100 mV, Cache 0 mV
223.9 W, 80°C


More of the same. Dropping only the core voltage another -50 mV resulted in a drop in power consumption and temperatures.

Core -150 mV, Cache 0 mV
200.1 W, 74°C


Another -50 mV less voltage for the core and a similar drop in power consumption and temperatures. At a core offset of -150 mV, this CPU has dropped 79W and 20°C.

Core -150 mV, Cache -100 mV
201.4 W, 74°C


For the final test I finally decided to increase the cache offset voltage to -100 mV. The results? Offsetting the cache made no difference to the CPU temperature and power consumption was basically the same. It went up 1 W which is within the margin of error.

At first I thought that maybe something was broken. To test for this, I adjusted only the cache voltage. I kept bumping the cache until finally the computer froze so the adjuster in ThrottleStop must be working.

My conclusion. For this 10th Gen CPU on this motherboard, the core and cache voltages seem to be 100% independent of each other. There is no need to adjust both voltages equally. Adjusting the cache offset makes no different in power consumption so for stability, best to leave it alone at +0.0000.

I did some more digging and it seems that on this board, the core and cache are being fed by 2 separate voltage regulators. I am not 100% sure about this so do not quote me on this. On my previous desktop board, the core and cache voltage regulators had the same ID number. Not so on this board. I think I might add a new window to TS that shows this information. It might be important info for people trying to figure out what they should be doing with the core and cache voltage so they can determine what works best on their board.

There was not power limit or temperature throttling during testing the Cinebench scores were fairly consistent.
Not that bad considering this is the 10900K's weak sister.

 

Hello, I'm trying to maximize power saving with my laptop, a Dell Precision 5550 with a i7-10850H.
I already disabled CFG Lock and enabled Overclock to make undervolt possibile.
Also about this I would like to report a suggestion. Newer Windows build enforce the use of Hyper-V and it could happen that TrottleStop works (opens) but doesn't really apply any offset. To fix this some features are required to be uninstalled or just disable hyper-v temporally (to test the undervolt and then set it permanent in the bios) using this command in the prompt with admin.

Code:

bcdedit /set hypervisorlaunchtype off

Could be useful to add a check since users can think that the tool doesn't work cause Intel locked it (even if you disabled the bios lock).

Anyway, I'm here to ask. How should I operate to find the right undervolt for the intel gpu and the system agent? Also does the new discoveries about the AVX voltage also applies to the 10th gen?
For now I have the following offset
CPU: -205
Cache: -110
iGPU and Uncore: -85
System Agent: -70
Tested with some game, 3-4 run with Cinebench and no error with the TS Bench.

Using a lower Cache immediately cause errors in TS Bench
With -100 System Agent cause system freeze (no blue screen just freeze)

Now with a quick search I can't really find any new info about iGPU and System Agent undervolt. Everybody say that they are not needed but the main goal here is lower cosumation as much as possibile so everything matters.

Does some one can confirm or better explain this and if this still applies to the 10th gen CPUs? Or some advice on how to test and improve the undervolt?

Thanks in advice for any help.

 

Cinebench R20 uses the AVX instructions. For some non-AVX testing, I decided to use the TS Bench test. It is limited to a maximum of 16 threads but that is good enough to create some heat and do some more voltage testing. Everything else was the same as above.

Core 0 mV, Cache 0 mV
212.7 W, 84°C


Core -50 mV, Cache 0 mV
191.1 W, 78°C


Core -100 mV, Cache 0 mV
171.9 W, 71°C


Pretty much the same as Cinebench AVX testing. Each step resulted in drops in power consumption and temperatures. When I tried running the TS Bench at a core offset of -150 mV, I ended up with a BSOD. A voltage setting that seemed stable during Cinebench R20 was not stable in the TS Bench test.

Core -125 mV, Cache 0 mV
163.5 W, 69°C


A core offset of -125 mV was stable. It shows a small improvement compared to -100 mV.

Core -130 mV, Cache 0 mV
162.0 W, 69°C


At -130 mV, the TS Bench was able to complete but there were 4 errors. This is the edge of stability. One might be game stable or Cinebench stable at -150 mV but long term, if the TS Bench is showing errors, you will probably end up corrupting Windows or some other files on your computer.

@Ansuel - Most people with a high performance gaming laptop have a dedicated GPU. The Intel GPU is not used or it is only lightly used. Undervolting the Intel GPU does not make a significant difference in power consumption. One user told me that undervolting his Intel GPU resulted in less stability and he was forced to reduce his CPU undervolt. It was not worth it so he set his Intel GPU back to zero.

Same thing applies to the System Agent. Does undervolting the System Agent decrease power consumption any significant amount? Most have decided to leave the Intel GPU, the iGPU Unslice and the System Agent at +0.0000. Have you found any tests that show an improvement when undervolting any of these three voltages?

Thanks for your bcdedit info.

 

Congrats on the new platform and thanks for the great testing (and thinking behind it) @unclewebb!

Can't wait to see what v10+ of ThrottleStop looks like and the capabilities it will have.

A question, can TS work on an MBP? I know the answer is no, but hopefully, I'm remembering wrong!

Thank you again for this great utility.