[SOLVED] Is Prime 95 Small FFT with AVX necessary to determine stability of 9900k OC?

[Norrat]

I've read that this test puts unrealistic workloads on the cpu and isn't suitable for a 9900k but there seems to be some debate about it still.

My overclocks pass 8h - Realbench (Half RAM selected/No WHEA errors), 8h Prime95 Blend (AVX Disabled), 4h - Memtest, 2h - Aida64 Stress FPU, 2h - Prime95 Small FFT (AVX Disabled) - is this enough?

 

[Phaaze88]

Prime95 small FFT with AVX covers cpu cooler thermal stability.
Asus Realbench 8hr or Cinebench R20 'infinite' loop cover voltage stability.
If you've passed both with no issues, then I'd say you're fine.

 

[Norrat]

Prime95 small FFT with AVX covers cpu cooler thermal stability.
Asus Realbench 8hr or Cinebench R20 'infinite' loop cover voltage stability.
If you've passed both with no issues, then I'd say you're fine.

Thanks for the info. Will have a go at P95 tomorrow.

 

[Phaaze88]

Thanks for the info. Will have a go at P95 tomorrow.

Careful though, as P95 with AVX on is rough for many coolers, so I suggest you disable AVX for the first couple of test runs.

 

[Norrat]

I've run 15min with AVX off a few times . My max temps were 74c, already pretty warm.

 

[Phaaze88]

Nice! Then see how it does with AVX on.

 

[Norrat]

It froze up when I tried running it with my 5GHz all cores OC, I got scared. But I've dropped it down to 4.9 now so here's to hoping

 

[Karadjgne]

AVX is an advanced vector instruction addition to x86. AVX itself is 256bit floating point calculations. AVX2 adds in 256bit integers. AVX-512 changes that from 256/256 to 512/512 on floating pount/integer.

AVX-512 is supported on only a few cpus, the 9900k being one of them.

Running Prime95 small fft is a 100% real world load. AVX, AVX2 and AVX-512 are BRUTAL on cpu loads and the consequential temps, driving a cpu up well beyond a 130% synthetic load.

Games do use AVX in small amounts to support physX partical calculations (think rock chunks in an explosion), I don't know of any that use AVX2 and most definitely not AVX-512. Those are mostly professional software design and engineering usage.

So kinda pointless for a homeowner pc to run in testing as a gaming load realistically won't see 100% and any small amounts of AVX get lost in the shuffle.

Test if you want to, but I'd disable AVX2 and AVX-512 at a minimum.

 

[Norrat]

AVX is an advanced vector instruction addition to x86. AVX itself is 256bit floating point calculations. AVX2 adds in 256bit integers. AVX-512 changes that from 256/256 to 512/512 on floating pount/integer.

AVX-512 is supported on only a few cpus, the 9900k being one of them.

Running Prime95 small fft is a 100% real world load. AVX, AVX2 and AVX-512 are BRUTAL on cpu loads and the consequential temps, driving a cpu up well beyond a 130% synthetic load.

Games do use AVX in small amounts to support physX partical calculations (think rock chunks in an explosion), I don't know of any that use AVX2 and most definitely not AVX-512. Those are mostly professional software design and engineering usage.

So kinda pointless for a homeowner pc to run in testing as a gaming load realistically won't see 100% and any small amounts of AVX get lost in the shuffle.

Test if you want to, but I'd disable AVX2 and AVX-512 at a minimum.

Thanks for the detailed response. Yep, I'm having no luck passing this with my 4.9 all cores OC. I get BSOD's or it just freezes up. I don't think this is due to thermals as my temps are maxing out around 80c so I up't the voltage to 1.3v which seems like a lot for 4.9 but the VDroop when running P95 drops it down to 1.145v. This is lowest I've seen it in any of these test. Don't really want to push the voltage higher or start messing will LLC again.

Besides the testing I've already completed, would you suggest anything else I need to run to verify stability? I don't want to send this ship out to sea until I'm sure.

 

[Karadjgne]

LLC should be left at auto unless it's not sufficient for the OC, at which point mid mid-high setting (3 out of 5 or turbo, not extreme) is the most you'll need.

Without it, the 9900k can have boot failures, failing to load windows because the cpu hits a voltage too low to continue.

View: https://youtu.be/H5AM7yhvsPU

Almost funny watching him get frustrated.

 

[Norrat]

LLC should be left at auto unless it's not sufficient for the OC, at which point mid mid-high setting (3 out of 5 or turbo, not extreme) is the most you'll need.

Without it, the 9900k can have boot failures, failing to load windows because the cpu hits a voltage too low to continue.

View: https://youtu.be/H5AM7yhvsPU

Almost funny watching him get frustrated.

Oh man that would have been frustrating. He handled it a lot better than I would.

Still don't see how this relates to my situation. His system wouldn't start when he went back to EVGA's stock settings which he'd never booted in on because he went straight for an OC. Doesn't mention LLC but does say the V was too low to boot, I'm getting in just fine and can run most anything except P95 with AVX.

Every guide I've read has said to set LLC to 5/6 out of 8 on Asus motherboards. I've been told if you go too high it can be unstable and there can also be undetectable voltage spikes that can damage your CPU.

Might have another go at it tomorrow.

 

[Karadjgne]

The point I was trying to make clear is that at stock settings, J2c LLC was at auto, so his actual voltages were not enough to actually boot with that cpu, but would pass post. By overclocking the cpu, he bypassed that and manually set the LLC, giving him enough to maintain stability. Your pc is doing similar in that vdroop isn't supplying enough when LLC is set to auto, so the OC is failing at 4.9GHz.

Asus is correct. A 5/6 out of 8 is medium to medium high, which is perfect for almost all OC attempts, 7/8 is honestly for the pros and world record OC.

LLC is an added voltage. The cpu loads change constantly, highs and lows. It's when the voltage is reaching a low, that's vdroop. The cpu instantly starts a load, and before the VRM's can bump up the voltage the cpu is suddenly out of juice, bsod. So LLC was adopted as a pre-emptive applied voltage so that the vdroop is covered for that instant. Pc remains stable.

Being an applied added voltage, it also affects the peak too, when the cpu has demanded high voltages from the VRM's and finally gotten supplied. So LLC will add to that voltage too and cumulative vcore goes through the roof. So you only need to set enough to cover the vdroop, not enough to swamp the cpu in power and putting undue stress on the VRM's.

So you'd drop vcore to maybe 1.2, add 5 LLC, which bumps the vdroop upto 1.2ish and vcore back to closer to 1.3v. (still set for 1.2v). Would give you a swing of @ 0.1v instead of the 0.15-0.20v you see now.

 

[Norrat]

Oh okay, thanks for the detailed response. I understand what you were getting at. My LLC is already set to 5 and I currently have a max swing of 0.155 - so maybe I should try for LLC6, although I have read a bunch of posts from people having problems with it,. Alternatively I suppose I could bump up the vcore to compensate for the droop.

Sorry if I was unclear. Still getting my head around all of this. Full settings

Vcore - 1.30v, 4.9GHz (all core), 0AVX Offset, Cache Ratio 45, LLC - 5, DRAM - 1.35v, VCCIO - 1.1v, VCCSA - 1.1v

 

[Karadjgne]

https://www.tweaktown.com/guides/9225/intel-core-i9-9900k-kf-overclocking-guide/amp.html

Your numbers look pretty much spot on, but there's always tweaks and differences between cpus and how they respond.

 

[Norrat]

Yep, maybe I didn't win the lottery with this one. Will keep tweaking it and see how I go. Thanks for all the advice! I'll check out that link.