[SOLVED] why do folks set an offset for AVX when OCing?

every guide i've seen suggest setting AVX offset at -2 or -3 when OCing. I know that will drop the OC by the same value when the CPU encounters workloads requiring AVX instruction sets, but
why do we want that offset?
 
Solution
Ralph,

“Stress” tests vary widely and can be characterized into two categories; stability tests which are fluctuating workloads, and thermal tests which are steady workloads. Intel tests their processors at a steady 100% TDP workload to validate Thermal Specifications. Prime95 version 26.6 Small FFT's is ideal for CPU thermal testing, because it's a steady 100% workload with steady Core temperatures which runs Hyperthreaded processors within +/- a few % of TDP. No other utility so closely replicates Intel's test conditions.

Utilities that don't overload or underload your processor will give you a valid thermal baseline. Here’s a comparison...

exroofer

Distinguished
Because AVX instructions sets, and software that use them, load the living crap out of the cpu.
And make a massive amount of heat.
So you need to dial back a wee bit, or it won't be stable, will take way too much voltage, and get even more stupidly hot than, say, a 9900k already gets.
Leading to thermal throttling.
 

Math Geek

Titan
Ambassador
AVX is pretty much the most taxing thing you can do to the cpu. so a heavy oc that is already pushing the cpu to the limits would be pushed that much further with an AVX workload. setting the offset will lower the oc a bit allowing the workload to still push the cpu but not to the point it starts hitting thermal or physical limits.

check out the recent reviews of new gen cpu's and you'll see what the AVX does to it. thermals go up quickly.
 

mjbn1977

Distinguished
Your normal overclock still will apply, but in the moment a core has AVX instructions to calculate, it will lower the clock rate according to the AVX offset only during those calculations. As soon as the core is getting other instructions it will go back to the normal OC rate.
 
Math Geek & exroofer: What you guys are saying just helped me understand something (i think). Right now, i haven't gone above 4.8 OC (waiting on a PSU cable to come in). But when i was burning in the new CPU, i ran P95 v26.6 small FFTs, so no AVX instruction sets. Temps were ranging 82-88 after 5 or so minutes of running, with two cores that would occasionally spike to 91 or 92C and core usage was showing 100%

Then, still without any AVX offset, i ran Handbrake compressing a large video file - it normally & predictably runs the CPU load or usage at 97-100%, but at 4.8 OC, it didn't go above 83-87% load max. Was that the CPU throttling? Temps were actually much cooler than when i ran P95, to the order of mid/high 70s to low 80s so it wasn't thermal throttling. I've got Vcore set at 1.240V and Fixed - think i'm going to run XTU to monitor and see what's happening while rendering
 

xxxlun4icexxx

Honorable
Jun 13, 2013
519
5
11,065




I'm not really sure why but setting the offset can help you get a higher multiplier as well. When I had my 7800x all cores were stable @ 4.4ghz but only if I had an offset of 4 (so 4.0 with avx). For some reason > 4ghz avx wasn't stable but regular work was. Ran it 24/7 for a few months with no issues.

Edit: Was your 9900k hitting 91/92c at stock frequencies? That's pretty hot. Fix: nevermind just re-read your reply.
 

Math Geek

Titan
Ambassador
prime 95 is a stronger work load than handbrake for sure. it pushes your cpu VERY hard which is why it is a good stability test. your temps on prime 95 are likely not going to happen in real life when gaming or normal workloads, so if your stable and temps are good in prime 95, then they should be good to go for anything. AVX is the only thing that actually taxes the cpu more. not a ton but more than prime 95 does. so if those temps are that high running prime 95, then you're gonna want better cooling to run heavy AVX apps.that or a big offset to make sure you don't start throttling.

remember every program uses a different amount of cores/threads, uses different instruction sets and uses the cpu to different degrees. so you may likely be fine in all normal uses but only see thermal problems in that small % of really taxing work like AVX.
 
From above: "Then, still without any AVX offset, i ran Handbrake compressing a large video file - it normally & predictably runs the CPU load or usage at 97-100%, but at 4.8 OC, it didn't go above 83-87% load max

would current throttling cause the cpu to reduce the load or usage?

My Vcore is set at 1.240V and FIxed (not adaptive) and I haven't explored how low i could take vcore, simply took that from MSI's guide, and figured i'd do a more in testing how low i could take it when that 2nd cable comes in. It did run stable thru three 1 hour burn-in runs, running P95
 

Math Geek

Titan
Ambassador
i am not fully versed in how handbrake uses the cpu overall. i do know that ram, hdd speed and other factors influence how it works. it's likely that at the higher speed your hitting a limiting factor from elsewhere so the program has to slow a touch to allow other parts of the pc to keep up.

back in the day when i used to do a lot of video work and such i always kept the temp files i was working with on a separate physical drive than where the finished file went. remember the hdd can only write or read at one time. not both. so compressing to the same drive means it has to read data and then write data in alternating fashion. going from one to another drive allows it to read fully from one and then write fully to the other. if i recall right i saw a good 50% increase in frames/second processed when i kept 2 drives working. this keeps the ram from filling up and having to wait for the hdd so it can offload the finished data dn then load new raw data and so on.

it's the old "weakest link" type scenario.
 

CompuTronix

Intel Master
Moderator
Ralph,

“Stress” tests vary widely and can be characterized into two categories; stability tests which are fluctuating workloads, and thermal tests which are steady workloads. Intel tests their processors at a steady 100% TDP workload to validate Thermal Specifications. Prime95 version 26.6 Small FFT's is ideal for CPU thermal testing, because it's a steady 100% workload with steady Core temperatures which runs Hyperthreaded processors within +/- a few % of TDP. No other utility so closely replicates Intel's test conditions.

Utilities that don't overload or underload your processor will give you a valid thermal baseline. Here’s a comparison of utilities grouped as thermal and stability tests according to % of TDP, averaged across six processor Generations at stock settings rounded to the nearest 5%:

Notice the extreme differences in workloads between Prime95 AVX versions such as 29.4 (upper left) compared to non-AVX 26.6.

All tests will show 100% CPU Utilization in Windows Task Manager, which indicates processor resource activity, not % TDP workload. Core temperatures correspond directly to Power dissipation (Watts) which is workload. Prime95 v26.6 Small FFT’s provides a true and steady 100% workload, so if Core temperatures are below 85°C, then your processor should run the most demanding real-world workloads without overheating.

Prime95 versions later than 26.6 run AVX/2 code on the CPU's Floating Point Unit (FPU), which is an unrealistic workload with Core temperatures up to 20°C higher than Prime95 v26.6. An AVX offset of -3 (300 MHz) or more may be needed to limit Core temperatures to 85°C. Even if users don’t run AVX apps, BIOS should still be configured for it, as certain utilities use AVX for stability testing.

AVX can be disabled in Prime95 versions later than 26.6 by inserting "CpuSupportsAVX=0" into the "local.txt" file in Prime95's folder. However, since Core temperatures will be the same as 26.6, it's easier to just use 26.6. As per Intel’s Datasheets, TDP and Thermal Specifications are validated “without AVX”.

See page 87, Thermal Considerations, Section 5.1.1, first paragraph, second sentence.

8th and 9th Generation Datasheet, Volume 1 - https://www.intel.com/content/www/us/en/products/docs/processors/core/8th-gen-core-family-datasheet-vol-1.html

CT :sol:
 
Solution
appreciate that, that tells me a lot, but what i still don't know (and probably won't until i get home this evening to test), with no AVX offset, what would have limited the cpu to 83-87% load, with the same predictably of the sun rising in the east, HandBrake (which employs AVX) is a core hog and runs all cores minimum 95-100% load, but usually 97-100%. I mean, i don't care if i'm rendering a low res low fps cartoon, it will still hit 97-100%

with my limited knowledge and picking up on what either exroofer or Math Geek said, could current throttling be limiting load? And again i have Vcore set to 1.240V and Fixed
 

CompuTronix

Intel Master
Moderator
The key here is understanding that "CPU utilization" only indicates processor resource activity, not % TDP workload. To visualize % TDP, you need to think in terms of power (Watts). Using Intel's intended stock settings, your 9900K should draw about 95 Watts + / - a few % as per specification, when tested with non-AVX Prime95 Small FFT's. However, due to variables such as VID, leakage current, ambient temperature and any non-stock cooling solution, actual TDP will vary somewhat.

When using any non-Intel custom BIOS settings such as MCE, manual or adaptive Vcore, LLC or various other overclock settings, under the same workload conditions (non-AVX Prime95 Small FFT's), power will exceed rated TDP, while CPU utilization never exceeds 100%. So to better understand what's happening with your CPU during any workloads, focus more on power dissipation rather than CPU utilization. This will allow you to determine at what power level (Watts) you may encounter problems, which in turn greatly aids in troubleshooting and analyzing performance.

Yes, that's entirely possible. Hardware Info is a great tool, which also provides graphs by clicking on a parameter such as CPU package power.

• Hardware Info shows advanced monitoring details - https://www.hwinfo.com/download.php

Select "Sensors Only" upon initial use.

CT :sol:
 


i can only assume h264 does employ avx, as clock ratio was set to 48, or 4.8 OC, i had set the avx offset to -2 and most cores were running 4622 (4.6) and every now and then one or two of the cores would bump up to 4.8, then drop back down to 4.6 - so i assume for short bursts they handled tasks that didn't include avx instruction sets
 
did some testing last night, and need to correct something i reported. I ran HandBrake both with and without any AVX offset, and saw the same cpu usage (79 to 88%) regardless of offset. Then i realized i just set this computer up and am on Windows 10 Pro for the first time - i'm thinking Windows 10 or the CPU / Z390 chipset are reigning in Handbrake's tendency to run at top speed. The earlier rig was a X99 chipset w/i7-5960X which was also 8C/16T on Win 7 Pro

But the nice part is, at it's reduced usage, temps are way down in the mid 60s to low 70s, which means i may not need to upgrade case and upgrade to liquid cooling.

Computronix thanks again for that detailed explaination
 

xxxlun4icexxx

Honorable
Jun 13, 2013
519
5
11,065


When you set all your cores to a 4.8 oc, do you still get the 5.0 turbo on the 2 cores?
 

xxxlun4icexxx

Honorable
Jun 13, 2013
519
5
11,065


so if I wanted to see for giggles, I could just disable the enhanced boost/turbo and set all cores to 5.0 and see if it's stable on auto voltage? Would that be the easiest way to test? I am just curious if I have a 5.0 capable chip or not. Would that be the easiest way to test? Do you recommend a specific avx offset?
 
from the general consensus, auto voltage will over volt your cpu unnecessarily and give you higher temps

any of the "auto" selections, including "Enhanced Turbo" seem to change settings i haven't identified, that;s why i assumed MSI set up Expert to give you a clean sheet to play with

I've got all the settings as per that guide, XMP, Turbo boost & Enhanced boost all off, ram is at 3200 with 1.35V
clock ratio is now at 4.9 with Vcore of 1.245.

I want to see the same but right now mine, when running P95 v26.6, small FFTs, at 4.9 I'm seeing temps in mid to high 80s with some spikes into low 90s - i;m still air cooled. I'm not going to explore 5.0 until i've installed a decent liquid cooling setup - been burning the web researching