OC i7-7700k to 5GHz, ASUS Z270G, everything seems stable but Prime95 ?

[gsrcrxsi]

Check my sig for full system specs. CPU was delidded with TG conductonaut between the die and IHS. MX4 between IHS and water block

I'll say that I haven't messed with overclocking since my old q6600/q9650/i7-920. So I've been out of the loop. but I started in with this chip, following this guide: http://edgeup.asus.com/2017/kaby-lake-overclocking-guide/

it was a great starting point, being that it focused on my same CPU, and most of the same BIOS options. My BIOS has an overwhelming amount of settings to tweak, making it hard to know what does what, and how useful it is for my purposes.

I want this thing to be 5GHz stable, I want it to run as cool as possible. I don't want to just hammer it with voltage.

So for now, I have a "mostly" stable setup it seems. I can run all my games and applications, and I don't see any problems with nearly all benchmarks or stress tests... EXCEPT Prime95. I dont know if its a version problem, or if prime95 isnt suited to my CPU architecture or something, but it just doesn't like my setup.

stress tests i've successfully run:
AIDA64 (CPU/cache) - 5+ hrs (max core temp, 69C)
LinX - 50runs, 5+hrs (max core temp, 86C)
Intel XTU - 5+hrs (max core temp, 70C)

Prime (small FFT) - usually fails after about 30mins. when it fails, it does not BSOD. It does not lock up. Prime95 simply reports that it got a calculation error, deems it a hardware failure, and drops a thread (50% core usage), usually after it drops one, it will drop more with more time. I'm not sure it's solely an AVX issue, as LinX uses AVX loads as well, uses a lot more RAM (like all of it), and even gets the cores slightly hotter than Prime, so what gives? I used to consider Prime the tell-all for stability, but now I'm questioning it since it's the ONLY thing that doesnt play nice.

Current BIOS settings: (if it's not listed, I left it on Auto, but there are A LOT of settings available)

Ai Tuner - XMP
XMP - DDR4 3200
BCLK - 100.00
AVX neg offset - 0
CPU ratio limit - 50
CPU cache ratio min - Auto
CPU cache ratio max - 45
CPU Load Line Cal - Level 6
VRM Spread Spectrum - Disabled
IA AC load line - 0.01
IA DC load line - 0.01
BCLK Adaptive Voltage - Enabled
CPU Core/Cache Voltage - Manual
Core Voltage Override - 1.320
DRAM Voltage - 1.353
CPU VCCIO Voltage - 1.100
CPU System Agent Voltage - 1.150

 

[SPL Tech]

Prime95 is unrealistically hard on the CPU. I can set my voltage low enough to fail Prime95 within seconds, yet I can game on it all day without a crash. Unless you're actually going to use Prime95 on your computer or use other hard AVX running software, dont worry about it. Run the computer at your current settings and if it never crashes then obviously there is no need to bump the voltage. If it crashes in game, bump the voltage a tad and try again.

 

[EarthDog]

Really, you don't ned to touch ANYTHING outside of vcore and multiplier.

If it isnt stable and you want it to be P95 stable, add more volts. Thats the game.

 

[gsrcrxsi]

i got it to run over 2.5hrs on p95 with more volts than i'd like, i think about 1.4v (and some other settings i dont remember now). but the CPU was pushing 87+C, even under water.

maybe i should just bite the bullet and try out some AVX neg offsets to drop the multi to 48-49 for AVX loads.

 

[trents]

You can also change the AVX negative offset. The reason you are failing P95 and not other kinds of stress testing is that newer versions of P95 run AVX 1 and AVX 2 instruction sets. AVX puts a lot more stress on your CPU but is not even used in general computing. Some specialized scientific applications do use it but common software does not. If you change the AVX offset so that the cores run a little slower if you should ever run something using AVX (like P95) then you should be stable.

 

[EarthDog]

i got it to run over 2.5hrs on p95 with more volts than i'd like, i think about 1.4v (and some other settings i dont remember now). but the CPU was pushing 87+C, even under water.

maybe i should just bite the bullet and try out some AVX neg offsets to drop the multi to 48-49 for AVX loads.

:golf clap:

Do you run anything which use AVX/AES instructions?

 

[gsrcrxsi]

You can also change the AVX negative offset. The reason you are failing P95 and not other kinds of stress testing is that newer versions of P95 run AVX 1 and AVX 2 instruction sets. AVX puts a lot more stress on your CPU but is not even used in general computing. Some specialized scientific applications do use it but common software does not. If you change the AVX offset so that the cores run a little slower if you should ever run something using AVX (like P95) then you should be stable.

LinX definitely uses AVX. i tried out the neg offset early on, and could see the core multi drop to 49 when running linx vs the 50 i had set (and the 50 it ran with other tests like aida64). LinX seem just as stressful in my opinion, core temps get a tad hotter than p95, but i did notice that the temps drop significantly during transition from run to run, so it seems the CPU is getting a small break every 5-10mins or so. So maybe that's a factor as well.

so i don't think its SOLEY because of AVX, but maybe the specific way p95 is using it?

:golf clap:

Do you run anything which use AVX/AES instructions?

not sure if any games are using AVX, do they? that's all I'm really doing with the machine right now. But I tend to try to "future proof" things by making it capable of any forseable task.

what kinds of applications, other than stress tests, usually use that kind of CPU instruction?

 

[trents]

Does LinX use AVX 1 and 2 or just 1?

Even so, my experience is that a machine that can't pass at least a 2 hr. P95 stress test, even if it seems stable otherwise, will often show instability later on.

 

[gsrcrxsi]

Im not sure if it uses AVX 1 or 2. maybe just 1. a google search didnt bring up any conclusive answers :/

 

[gsrcrxsi]

ok so here's what i did.

left cpu voltage the same. 1.320v
increased the core multi to 51 (5.1GHz)
applied AVX offset of -2 (4.9GHz)

ran a 2hr small FFT without issue, and an 8hr blend test without issue, correctly running at the 49 multi for 4.9GHz.

now i need to figure how to force Prime to run without AVX to test the 5.1 speed.

 

[mackerel]

You can edit the P95 config file to disable or configure some CPU features. I think what you need to edit is mentioned in the readme file that comes with it.

 

[EarthDog]

Or use an older version without it.

 

[gsrcrxsi]

Yup. Google search pulled up how to do it. Had to add some lines to the local.txt file.

Started the 5.1 non-AVX test, will let it run all day and check it after work.

So far,
4.9 GHz Prime w/ AVX, max temp 81C
4.9 GHz LinX w/ AVX, max temp 85C
4.9 GHz AIDA64 w/ AVX (CPU/cache checked), max temp 65C
5.1 GHz Prime no-AVX, max temp 75C
5.1 GHz AIDA64 no-AVX (CPU only), max temp 60C

- - - Updated - - -

I wonder how far I can push the non-AVX clocks on this thing...

 

[gsrcrxsi]

well i guess 5.1 wasnt great. came home to a restarted computer. logs show that it restarted within minutes of me leaving the house lol.

played around with it some more, and it was just wanting too much voltage IMO. i didn't want to go over 1.35 really.

backed the settings to 50 multi, 1.325 vcore, -1 AVX offset.

re-running prime95 26.6 (no-AVX)now, then i'll re-run prime95 v29.3 (latest) on the same 4.9 settings as before. should pass i guess.

 

[batboy]

This is part of the reason I don't use Prime95. While I'm not impressed with the XTU stress test, the XTU benchmark is really a good, quick test of stability. I'm not saying one run means you're 100% stable, just saying if it don't pass XTU, then it most certainly is not stable. What is truly stable? Depends on what you do. If you have to calculate mission critical data and lives depend on it, then maybe I'd worry about Prime95. Otherwise, no worries.

I looked over your settings from the first post and don't see much I'd change, but if I were you, I might try a tad bit more DIMM voltage to make sure the memory is stable. I think 5.0 is a good point to be long term. You just need some minor tweaking. I'd try using your same settings (listed in post #1) except use AVX negative offset 1. CPU cache ratio max I would back down to 44 (that's what I always use for 50X). I would try a minor voltage increase to vcore and DIMM, maybe try 1.325v for core voltage and bump the DRAM Voltage 1.36v. If you use everything else the same and try those changes, I bet you'll pass Prime95.

Seems like all the OC guides say set IA AC load line to 0.01v and IA DC load line to 0.01v. Nobody ever says why do this. I have tried with these settings at default and with 0.01. I have yet to notice any difference, so I went back to the default settings. Leaving it the way you have it should be no problem, though.

There is one setting you listed that I have no idea what it is: CPU Load Line Cal - Level 6. That's a setting I've never touched. What is default?

 

[EarthDog]

Its LLC batboy... prevents vdroop.

 

[gsrcrxsi]

This is part of the reason I don't use Prime95. While I'm not impressed with the XTU stress test, the XTU benchmark is really a good, quick test of stability. I'm not saying one run means you're 100% stable, just saying if it don't pass XTU, then it most certainly is not stable. What is truly stable? Depends on what you do. If you have to calculate mission critical data and lives depend on it, then maybe I'd worry about Prime95. Otherwise, no worries.

I looked over your settings from the first post and don't see much I'd change, but if I were you, I might try a tad bit more DIMM voltage to make sure the memory is stable. I think 5.0 is a good point to be long term. You just need some minor tweaking. I'd try using your same settings (listed in post #1) except use AVX negative offset 1. CPU cache ratio max I would back down to 44 (that's what I always use for 50X). I would try a minor voltage increase to vcore and DIMM, maybe try 1.325v for core voltage and bump the DRAM Voltage 1.36v. If you use everything else the same and try those changes, I bet you'll pass Prime95.

Seems like all the OC guides say set IA AC load line to 0.01v and IA DC load line to 0.01v. Nobody ever says why do this. I have tried with these settings at default and with 0.01. I have yet to notice any difference, so I went back to the default settings. Leaving it the way you have it should be no problem, though.

There is one setting you listed that I have no idea what it is: CPU Load Line Cal - Level 6. That's a setting I've never touched. What is default?

not mission critical, just me being a perfectionist lol. If it was truly a mission critical system, I wouldn't OC it at all.

i'll see about bumping DRAM to 1.36 if i need to. maybe that will allow me to drop Vcore slightly if thats the source of instability. So far i ran an 8+hr blend test of Prime95 v26.6 (no-AXV) @ 5.0GHz. thats a pass for me. It's running a blend test of Prime95 v29.3 (AVX/FMA3). if it makes it >8hrs I'll call that a pass too and be done with it. I've currently got Vcore set to 1.325.

I agree that the guides say to put IA LL AC/DC to 0.01, but no reasoning. wish i knew why. but to correct you a tad, at least according to the description in the BIOS, the setting is not in volts, but rather in 1/100 mOhm, this is a resistance. the parameter has a range of 0-6249 (0-62.49 mOhm).


So a setting of 0.01 should be 0.0001 mOhm (or 0.1 µOhm). Some online posts say that this will have some kind of effect on the vDroop when using adaptive voltage, but as for "HOW", that's a mystery to me. Oh well, if it ain't broke don't fix it. Auto probably works fine here too.

And yes, CPU LLC, ASUS has 7 levels, 1 being the most vDroop, 7 being the least vDroop. I'm running on the higher end of the scale. But i found with anything 5 or lower it was very unstable (at least when i was trying to run AVX Prime @ 5GHz, no AVX offset)

 

[batboy]

I stand corrected and have learned a little myself. Good luck, sounds like you're about there, just a bit of final tweaking. Once you get the CPU rock stable, maybe you can play with the RAM a little? What frequency are you running? I presume 3200? With a bit more voltage, I was able to run the same tight timings of my DDR4-3600 at 3733 MHz which helps in many of the benchmarks and apps.

 

[gsrcrxsi]

I left prime v29.3 running all day, 17+ hrs lol. no failures and no errors. so it's stable now. maybe at a later date i'll try to mess with it to get 5.1 out of the non-AVX clocks, but for now i'm not gonna mess with it anymore.

memory is at the stock XMP profile. DDR4-3200MHz, 1.35v, 16-16-16-36.

 

[gsrcrxsi]

I ran LinX v0.6.8 on this config just for the hell of it. since it uses AVX, it downclocks to 49 multi.

still hit max core temp of 90C. and this is a Delidded CPU on water cooling. crazy.