AVX power limit for stability

Associate
Joined
30 Aug 2018
Posts
2,483
So often i hear that p95 AVX is unrealistic (unless you are searching for primes ofc) and a power virus. This is particularly true for overclocking and overclocking stability tests where the power demands and heat of p95 AVX generates instability.

I've seen a couple of approaches to dealing with this instability namely:
  1. Using an AVX offset.
  2. Not testing with p95 AVX
1 is great for dealing with Prime, but when it comes to less demanding AVX workloads such as gaming for example you find that the inclusion of AVX means your overclock is limited to your AVX offset. Not ideal.

2 is a valid option, but it also begs the question who gets to decide what is and what isn't stable and how can you say your CPU is stable if you don't test for stability in all things?

I personally don't think being stable in P95 AVX is necessary due to the unrealistic nature of the test for everyday workloads. If you chase AVX stability in P95 you end up using such high vcore for everything else that you end up artificially limiting your oc through voltage or heat output anyway. An AVX offset that allows you to run prime 95 AVX is also going to be aggressive compared to the kind of offset needed for regular AVX loads.

Ideally there would be a way to overclock the CPU for non AVX workloads while retaining stability in the kinds of AVX workloads you see in day to day activities. If you can do this AND find stability when running prime then that would be, imho, the perfect overclock.

So i want regular software to run at my max overclock of say 5.4GHz while also allowing reasonable AVX loads like BF5 and Realbench to run AVX at 5.4GHz. I also want to know that i wont run into a situation where AVX use becomes so high that it has the potential to introduce instability.

I was browsing reddit and I think I have found exactly what I am looking for. I can't test it right now because I'm waiting on a replacement die frame but I intend to try it as soon as I'm able.

https://www.reddit.com/r/overclocking/comments/a7ef1q/using_power_limits_for_avx_stability_on/

In the reddit post the person overclocks for stability and then controls AVX power consumption and heat output by setting power limits. This looks like the solution to being able to run AVX instructions in BF5 and realbench at max OC frequency, while also keeping P95 AVX and other power virus loads stable.

Yes the draw back is that P95 won't score as high or run as fast compared to an unlimited overclock, but so long as the limits imposed don't affect any other software then you are trading speed in P95 for high speed in anything less demaning, while ensuring absolute stability in every workload.

Is this crazy? It seems like a have your cake and eat it solution.
 
Im at 5 with a 0 offset avx prime stable chip can do 5.3 with a 1 offset but higher voltage im happy with 5 bloody good chip

What im trying to say is even at 5 with say a 2 offset u not going to see much difference in games and programmes just run prime avx for an hour if its stable your good to go for most day to day use

By the way whats currently in your rig can give u better advice then
 
I would aim for a 24/7 overclock with enough voltage to compensate for the AVX instruction, without using an offset. That way you know exactly how much voltage you need for AVX stability.

Test it 4.7-5Ghz say, AVX, find stability and compare temps and volts. When temps hit 80 ish call it quits. This will be your max AVX clock/volts and will later help you determine the offset if you decide to clock higher for non AVX tasks (I wouldn’t bother for the difference 200mhz or so will make)
 
https://www.reddit.com/r/overclocking/comments/a7ef1q/using_power_limits_for_avx_stability_on/

This sounds just like what I was after. His aim was the same, to be able to squeeze 300MHz more than otherwise for daily usage. With AVX becoming a bit more common, no longer restricted to some benchmarks, the old advice to set Adaptive just don't run AVX benchmarks doesn't work so well anymore.

My chip needs 1.35v for 4.5GHz. Having it on Fixed voltage 24/7 is not something I want. And on Adaptive, the Vcore will spike to 1.45v with AVX which is something I really don't want.

Am about to give it a go, just thought I'd share for others with older chips who've wondered about this.
 
Just to report back on this - it does everything I wanted except it doesn't prevent Vcore from reaching 1.45v on non-full load AVX. i.e. It controls the Vcore very well on full load AVX, and obviously power draw and heat (if those happen to be concerns), but when gaming, because the current limit and power limit aren't being maxed, there are still some Vcore spikes to 1.45v. Presumably due to brief moments of AVX in use (since the same is not experienced with Fixed voltage as opposed to Adaptive).

So it'll be helpful in some scenarios (power and heat), but not if one aim is to stop Vcore exceeding a certain level. For that, the only way still seems to be Fixed voltage not Adaptive. Or lower Vcore enough to prevent it, e.g. 1.3v to keep it from going past 1.4v, but then obviously that can impact the overclock.
 
Just came across this thread. Posted a thread about this yesterday without realizing there was one already: https://forums.overclockers.co.uk/t...ity-on-platforms-with-no-avx-offset.18850351/

It's really great, other than being unable to control AVX Vcore spikes when using Adaptive. But having decided I can't have absolutely all the cake, and going with Fixed, it's helped my 24/7 overclock in this respect:

4.5GHz @ 1.35v is good for just about everything. But bluescreens with H.265/HEVC encoding. Options are lower overclock (don't want), increase Vcore (don't want), or implement some beneficial current and power limits. After finding how much current and power Cinebench R15 (non-AVX) uses, I set current and power limits just beyond that. Ran Cinebench R20 (AVX) and saw the frequency clock to 4.2GHz and stay there the whole run, in this instance just like would happen with a -3 AVX Offset on modern CPUs/boards. More importantly, when running the H.265 encode, it didn't bluescreen, it just flipped through 4.5, 4.4, 4.3 and 4.2 whenever the current and power limitations kicked in. So it's really useful.

Also, its even better than an AVX Offset. Better, not easier, as you have to go through the process of identifying the sweet spot for current and power. Why? Because say you set a -1 AVX Offset, and it's stable for most of a task, but bluescreens in one place. Or you set a -2 AVX Offset, and the same. You set a -3, and now it doesn't bluescreen, but you wonder wouldn't it be nice if it was able to flick through speeds as and when needed instead of always being 300MHz lower. And that's exactly what this AVX power limit method does.
 
Last edited:
This is what i am hoping to do eventually. It avoids an offset and can be used to keep extreme avx loads in check via power limits without limiting reasonable avx loads like gaming (or anything not p95 :D ). Glad you got it working and found it worthwhile.
 
This is what i am hoping to do eventually. It avoids an offset and can be used to keep extreme avx loads in check via power limits without limiting reasonable avx loads like gaming (or anything not p95 :D ). Glad you got it working and found it worthwhile.

Thanks. Me too.

I was thinking boy it'll be nice when I get a new board in the future with AVX Offset. Now I know I'll use this instead of that anyway lol. I'll post in the other thread to ask the Mods to close it, to have just this one up.
 
Avx offsets can actually increase the vcore needed for a given overclock which is something I wasn't immediately aware of. Limiting AVX power draw in this way also eliminates the possibility of that happening so as long as it works as intended, it is preferable to an avx offset. Win win. \o/
 
Have you tried temperature offset to start throttling at a lower temp rather than power throttling, say setting TCC activation to 85C?
 
Back
Top Bottom