Undervolting lowers CPU performance

Just tried -110mv UV and faced nasty results.

Pros: temps lowered by 10-15 celsius
Cons: FPS dropped dramatically (~10fps) in CPU demanding games like SC2, Heroes of the storm.

Going to try with -75mv to see a difference.

 

Undervolting alone shouldn't cause a decrease in performance, to my knowledge. If anything, in some cases it can increase performance by allowing a CPU to stay at boost clocks for longer before thermal throttling.

Any undervolt too severe usually results in a crash, not decreased performance.

You should monitor CPU clocks and package power to see if you're not boosting or if you're TDP limited. Are you using XTU to undervolt?

It would probably also help if you post your laptop model / specs.

 

I am using Throttlestop 8.6, forgot to mention, my CPU is 7700HQ / 1060 Max-Q (Inspiron 7577), also speed shift is enabled in throttlestop.

Just noticed, that overwatch for example, stays on the same fps lvl 60-70, with or without uv, doesn't matter. But HOTS getting lower fps...

 

Not sure you are interpreting the data correctly. A 10-15 degree drop in CPU temp is significant. Especially, if it stays that way on average. A 10 fps drop is meaningless unless you are getting that through a synthetic benchmark. Games can be great for checking stability,but the random nature of game playing makes them lousy for obtaining consistent benchmark results.

See this video for example and why the author chose the specific area of the game to play in:

 

You mean it's my subjective feeling about losing fps?

 

Not what I said or intended. 10 fps drop may not be statistically significant especially if you are not controlling random variables.

 

Well, i didn't test on exact same map, with exact effects. Perhaps, 7700hq is not enough for heroes of the storm?

 

Use benchmarking applications for consistent loads. Keep in mind that core performance is impacted by temperatire. Your GPU could also be providing enough heat to the chassis to cause thermal throttling.

Are you using HWINFO or AIDA 64? Post screenshots of the results with and without undervolting using the same benchmark load.

 

Not using both of them. Should i try AIDA 64 benchmarks? Would it help me measure cpu performance? Also synthetic benchmarks is kinda not like the games, would it be accurate?

 

Download HWINFO64. It is free, and will help you get familiar with the sensor information your machine is telling you. AIDA64 is more advanced and requires a payment to unlock all features.

Once you have HWINFO64, peruse the Throttlestop forum and others here where many folks have had similar concerns as you. Including me.

 

Basically just bench before/after UV, right?

 

Try this guide. One among Avogadro's number available online:

http://forum.notebookreview.com/threads/the-undervolting-guide.235824/

 

Another example from the Alienware 17 R5 thread courtesy of @PandahNZ

http://forum.notebookreview.com/thr...e-guide-results.797373/page-358#post-10789474

 

btw, is 3DMark legit source for testing?

 

3DMark has a whole series of tests you can run. The endpoint is to get your cpu and gpu sufficiently stressed so that you can test for stability.

Some tests out there you should stay away from since they are specifically enthusiast-focused and could irreparably stress your components. MSI Kombustor for example.

 

I've never heard or seen any case where undervolting the CPU lower performance if it's stable.

Undervolting the GPU however can yield less frames, the GTX 1050 on my G3 is a good example: anything under 850mv @ 1493MHz can lower the performance by up to 10% at 800mv but with on my nephew's GTX 1060 Max-Q: 800mv or 850mv @ 1607Mhz gets the exact same performance only with lower temps at 800mv.

 

Well, how would you explain performance drop with your 1050 Ti then? Not enough voltage and its downclocking due to less energy.

 

The only explanation I can give since both voltages are stable @ 1493Mhz and does not downclock according to GPU-Z and HWMonitor is at 800mv the card is not allowed the efficient TDP needed for proper performance @ 1493Mhz, while at 850mv with the same clocks the TDP is raised slightly probably 5 watts to allow for the proper performance @ 1493Mhz (not really a performance gain since at default voltage curve the card runs at 900mv @ 1493Mhz, more like undervolting too much squeezes the TDP).

And since the voltage is already so low at under 850mv, the temp difference between 800mv and 850mv @ 1493Mhz is negligible being only 2-3 degree lower on 800mv but the 10% performance gain by just upping the voltage to 850mv with the same clocks is pretty big and more than worth it to keep there because the GTX 1050 runs very cool.

 

Why you think the same logic won't apply for cpu undervolting?

 

The same logic applies, except the CPU usually becomes unstable before it loses performance, and so will the GPU - but it shows up as lower performance rather than a hang, driver freeze / crash, or artifacts in the image.

The instability in CPU undervolt being to much (voltage too low) includes having prime95 threads die with math errors, lower performance - result rejected due to bad computation.

The thought that you can say undervolting causes performance loss is the same as saying that undervolting causes instability and math errors, it does if you undervolt too far.

Undervolting is a process, one value of undervolt isn't as good as another, you are tuning through a series of reductions in voltage - increases in undervolt value (-minus value) until you hit: instability, computation errors, less performance - then you reduce the undervolt and test again.

Once you have an undervolt value that is stable with no undesirable side effects at 100% load - like instability, computation errors, or less performance - then you tune for idle.

Idle undervolt may be +5mV or +10mV reduction in undervolt, say from -150mV to -145mV or -140mV.

When a laptop is idle the CPU will ratchet down the voltage to save power until it reaches a minimum setting, which may be too low with the undervolt that worked at 100% load... and it BSOD's or reboots.

It's helpful to exit all programs, stop all background processes - like Netflix, bitorrent, or any other process that might put load on the CPU, you want the laptop to sit for 30 minutes to 1 hour as idle as possible to pass.

So, in short, please stop spreading FUD like "undervolting reduces performance" without adding the word "bad" or "incomplete" in front of it, because the end result of the successful completion of the process of undervolting is stable operation with full performance and a cooler running CPU.

Quite often tuning the undervolt increases performance. If you were thermal throttling and undervolting stops the thermal throttling, that improves performance. If you were power throttling then reducing the voltage can also improve performance.

With higher core count CPU's and inadequate cooling undervolting usually does both of these things, giving you better performance than the out of the box BIOS CPU voltage tune.

The too high voltage to the CPU as a standard BIOS setting has been the norm for many years. That's why undervolting is so popular. The laptop maker isn't tuning the voltage for every laptop, but that is what is needed to match the voltage to each unique CPU. You are completing the tuning process the vendor should be doing for us by undervolting.

Undervolting is a good thing. Remember that while you are tuning for the optimal negative CPU voltage offset, or undervolt for your laptop's unique CPU.

Any questions?

 

Well, thanks for the extended answer. What you can recommend as a testing tool for cpu performance measuring before / after UV? I'm stable at -120mv, but got the feeling, that cpu performing worse in game (losing fps)

 

We've all been there, perception is 90% of reality, even if you are going into a tougher area in the game and seeing lower FPS due to the game load - it still feels attributable to the tuning you've just done - whether that is undervolting or OC settings on the CPU / GPU, too many changes cause confusion.

That's why I suggest living on each tuning change until you are happy with it, and then have a good base in factual measurements from benchmarks either in game benchmark tools or external benchmark tools.

If you don't have a repeatable test that within the normal % of error can get the same range of result - like within 5% range higher / lower, and you are "eye-balling" it freestyle with random game play, you'll chase your tail for days.

Also remember that Windows 10 is a hell-hole of background telemetry collection, report generation for that data, and uploading it to who knows where many times a day. Before Windows 10 started doing this we could pretty much tune out most of the background processes and get more repeatable test numbers. It's still possible but requires more research and effort.

Running tests several times and averaging them, even plotting them on a graph, gives you an idea of whether the results are hovering around a straight line average or if you have wild results mixed in - like too low results due to something firing off in the background.

I've been doing this for a long time, and after a few reports from owners of new CPU's there begins to form a range of undervolt and you can pick a center value to start from. Usually -100mV is a good place to start, and even end undervolting, as for most people getting the -10C reduction in temperature at 100% CPU load with -100mV undervolt is enough to stop thermal throttling; so there's no need to spend time to tune for a few more mV improvement. Unless you want to.

There are more reports from each succeeding generation of CPU of higher undervolts, with -150mV or more reported for the 8th gen CPU's. It's not a contest though, you are looking for the "right" value for your CPU, and if that's -120mV then that's what you want to end on.

When I am tuning an unknown CPU I will use the XTU tool Stress Test or XTU Benchmark when doing the first couple of tests, as usually that is enough to fail a bad undervolt if I went to high - I like to quickly get to the right value so I start at -100mV and jump up -10mV or -20mV at a time until I BSOD - and then back off that last increase plus +5mV for the next test.

For the final testing I run prime95 small fft with AVX/FMA instructions disabled:

Exit prime95 and add these lines to the top of local.txt:
CPUSupportsAVX=0
CpuSupportsAVX2=0
CpuSupportsAVX512F = 0
CPUSupportsFMA3=0
CpuSupportsFMA4 = 0

Then reset the readings in hwinfo64 and restart prime95 small FFT's.

You are using hwinfo64 to monitor the CPU temperature, power usage, and voltages?

Watch for prime95 threads to exit, and if they do even if you seem stable, reduce the undervolt +5mV and run prime95 again - make sure all the threads stay running. 15 minutes is usually enough for the last test, 5 minutes for the pre-testing attempts. There's no need to run prime95 longer, like for hours overnight - it's just stressing the components for no useful purpose.

Please come back and let us know how it goes, and what your final Idle stable undervolt is for your CPU.

 

I'm using HWMonitor to monitor temps/frequencies etc. So, generally your point is UV is not decreasing performance, it just BSOD's, when there's not enough voltage, right?

 

That's why I use prime95, if the threads exit but the OS still hasn't BSOD'd, I know the undervolt is too large - not enough voltage.

I back off +5mV test again, and usually that is enough to allow prime95 threads to keep running. If not I back off another +5mV.

I want to find the stability point quickly, and prime95 small fft with AVX / FMA disabled finds instability quickly.

After you have a stable 100% load and stable idle undervolt setting, your laptop shouldn't thermal throttle, and your performance is optimized for use.

Before Pascal, tuning the GPU you can make settings for the OC of core and memory that will be stable but have less performance at a given voltage than lower OC settings. Tuning the GPU involves more than just undervolting at stock settins, adding voltage might help at higher OC - as you need more voltage for higher OC's if allowed on some GPU's.

The tuning is different for Pascal than previous generations, less voltage can give better frequency response and performance vs manually increasing the core and memory clocks.

We don't know what it will be like with the Nvidia 2xxx GPU's - the RT and Tensor core tuning - voltage usage - may enter into getting the best rasterization performance.

Also, if you are going to OC the CPU past stock settings, your undervolt needs to be reduced as the OC increases. With the 7th and 8th gen you might reach the OC maximum for the CPU before you run out of undervolt headroom.

On previous CPU generations we had to add overvolt after a certain OC level. For me the 5950HQ @ 4.3ghz needed +100mV voltage offset - an overvolt rather than an undervolt. At 4.0ghz no undervolt or overvolt was possible / needed - that is the BIOS CPU voltage balance point.

So, each time you change the multipliers on the CPU - if your CPU allows it - you can retune the CPU voltage to match. For lower multipliers you can increase the undervolt, for higher multipliers reduce the undervolt.

When you have tuned for maximum stable performance for the CPU, set it back to defaults and do the tuning for the GPU with the CPU at stock settings. Isolating the tuning helps find the maximum for each individually.

Not all laptops have the power and cooling headroom to run both the CPU and GPU at full power, so pick one to set to the maximum and detune the OC of the other a little a bit - usually not a lot - for stable operation with both.

 

Steps to test with prime95:

1. Using your settings on top of local.txt
2. Then run prime95 (?) what test to choose? Where to monitor threads? (if it's exited)

Thank you

 

small fft...

The threads each have a little dialog in the monitoring window, it should be obvious when one stops - the text stops scrolling in that little window, and you can scroll through the list from time to time to check them all, or grow the window to large enough to see all of them running - and at a glance you should see a window where the text stopped scrolling by.

 

Okay, gonna try it, when i'm at home, thank you, bro