MCE explanations and others

[Luck100]

Well, I guess just use Intel's own version then.


Ahh ok, thanks, forgot about that.

What exact settings are you running on the Hero to achieve the balance of AVX/non-AVX?

Not sure what you mean by balance of AVX/non-AVX. These are the setting which I've changed from default values:

Ai Overclock Tuner [XMP I]
XMP [XMP DDR4-3200 16-18-18-36-1.35V]
BCLK Frequency [100.0000]
ASUS MultiCore Enhancement [Disabled]
AVX Instruction Core Ratio Negative Offset [0]
CPU Core Ratio [Sync All Cores]
1-Core Ratio Limit [49]
DRAM CAS# Latency [16]
DRAM RAS# to CAS# Delay [18]
DRAM RAS# ACT Time [36]
CPU Load-line Calibration [Level 6]
Long Duration Package Power Limit [180]
Package Power Time Window [5]
Short Duration Package Power Limit [225]
CPU Core/Cache Current Limit Max. [255.75]
CPU Core/Cache Voltage [Offset Mode]
- Offset Mode Sign [+]
- CPU Core Voltage Offset [0.035]
DRAM Voltage [1.3500]
CPU VCCIO Voltage [1.15000]
CPU System Agent Voltage [1.10000]

[vvoid]

Ok, short follow-up, I've tested power limiting AVX/FMA3 load a bit. I see 2 problems with it:

1. The transition from the high multiplier towards the lower, power-throttled one is still occuring (well, of course, what was I thinking?). It might still be triggering the transient crash and nothing's really gained.

2. The limited frequency probably is whithin "normal" operating range (36x - 49x for my 9700k), applying adaptive voltage VIDs as it's supposed to. But this might not be enough for the torturing AVX load, leading again to a crash. Way out is applying a positive offset voltage, but then, this means higher voltages overall.

So... I don't really know. Might be that in the end we actually end out roughly in the same voltage regions as when using AVX offset in the first place. Nevertheless, just for "safety" reasons it might be a good idea to limit power consumption anyway.

Oh, and why actually use PL1/PL2? Limiting IccMax achieves the same goal and is much simpler to test/stabilize, or am I missing something?

[Luck100]

Ok, short follow-up, I've tested power limiting AVX/FMA3 load a bit. I see 2 problems with it:

1. The transition from the high multiplier towards the lower, power-throttled one is still occuring (well, of course, what was I thinking?). It might still be triggering the transient crash and nothing's really gained.

2. The limited frequency probably is whithin "normal" operating range (36x - 49x for my 9700k), applying adaptive voltage VIDs as it's supposed to. But this might not be enough for the torturing AVX load, leading again to a crash. Way out is applying a positive offset voltage, but then, this means higher voltages overall.

So... I don't really know. Might be that in the end we actually end out roughly in the same voltage regions as when using AVX offset in the first place. Nevertheless, just for "safety" reasons it might be a good idea to limit power consumption anyway.

Oh, and why actually use PL1/PL2? Limiting IccMax achieves the same goal and is much simpler to test/stabilize, or am I missing something?

I'm not using power limiting for stability. My chosen voltage/clock is stable. But I'll hit 100C if I run a heavy AVX load for very long. The power limiting just knocks the clock/voltage down gracefully for those cases so I don't go bouncing off the 100C thermal limit (which will cause a much bigger downclock anyway, aside from any concerns about being too hot). I think this is better than an AVX offset because I can still run light AVX loads like BF5 without any downclocking at all (because they don't draw too much power). Bottom line, this won't help if you're not AVX stable already.

I suppose I could use IccMax instead, but I never tried it. There's nothing complicated about setting PL1/PL2 and you can allow a higher short-term power burst with PL2.

[vvoid]

Ahh ok sorry, I might have misunderstood you then. My goal was to exploit power-limiting for stabilizing AVX offset transitions - or really any kind of clock transitions - in some way, and in that sense, I think it doesn't really work. For generally keeping temperatures in check, sure.

That said, I wonder how you can really be sure your "chosen voltage/clock" is truly stable if you cannot run prolonged, full AVX-load because of temps?
EDIT: I guess forget that question. You're simply stable, enough vcore applied, but you would just need better cooling. All clear.

Anyway, after testing some more, I actually do think limiting power is a good idea and the better/easier route compared to using AVX-offset. It might not be 100% Prime-stable in the end, but in practice this doesn't matter, at least in my use cases. Failing Prime still is far better than a random BSOD, which is prone to happen when using AVX-offset with vcore too low...

[vvoid]

Hmm, not much interest in this thread it seems, yet this is a very essential topic for CPU overclocking imo, esp. the AVX-offset thing.
Some more testing on my part shows that XI Gene performs similar in this respect. Guardband appears to be even larger than on the Z390-F, but it's not directly comparable because the Gene has 8 LLC levels (use 6 or 5). Overall Gene is better, that's for sure, albeit not by that much, at least for normal, non-subambient operation.

One last recommendation in general: Don't use AVX-Offset! It's only leading to instabilities, requiring more vcore in the end than with limitng IccMax instead. The trick with IccMax is to apply positive offset voltage when using adaptive voltage, that's all. Without that it will crash because of the too low AVX-voltage on lower multipliers. For fixed voltage I don't know and I wouldn't suggest to use that anyway. Think of power usage and I've yet to see cases where higher 24/7 overclocks have been achieved using fixed voltage in contrast to to adaptive voltage, this is a myth imo. Adaptive Voltage is the way to go since Skylake. Fixed is easier to stabilize, yes, but it's not the correct route for a 24/7 overclock.

[Stu @ MSD]

Wow, the standard of knowledge around here has improved somewhat since I left all those years ago!

Great topic guys, I am looking forward to learning from you.

[RobertoSampaio]

My CPU is running 5.6GHz with a 280mm AIO.

Motherboard: ASUS Maximus XII Formula - BIOS 2103 - SVID@Trained (AC/DC loadline: 0.510/1.100mOhm) - LLC#4@900KHz - Adaptive@1.524V - VCCSA@1.240v - VCCIO@1.190v

CPU: Intel Core i9-10900KF SP86 => 56x4 - 55x6 - 54x8 - 53x10 <= (+2Boost OCTVB profile, VMaxStress, VoltageOptimization) - Ring@8~47x -
Full Load: 51x@1.240v

V/F Offsets: VF#1@659-8mv / VF#2@794-3mv / VF#3@914-3mv / VF#4@1054-3mv / VF#5@1169-18mv / VF#6@1328-12mv / VF#7@1328+53mv / VF#8@1428+95mv

[vvoid]

Well, congrats on this chip, seems to oc very well, but also very high vcore, no?
I wouldn't recommend running that 24/7, far too high. Remember, Maximus boards are about correct on voltage, in contrast to the lower tier boards where you can substract at least ~100mV from the HWInfo reading under full load. Real 1.47V is far too high and I doubt you can keep temps in check with an AIO. Show screenshot under full, prolonged Prime95 small-FFT non-AVX load, that's when we're talking!

[RobertoSampaio]

Well, congrats on this chip, seems to oc very well, but also very high vcore, no?
I wouldn't recommend running that 24/7, far too high. Remember, Maximus boards are about correct on voltage, in contrast to the lower tier boards where you can substract at least ~100mV from the HWInfo reading under full load. Real 1.47V is far too high and I doubt you can keep temps in check with an AIO. Show screenshot under full, prolonged Prime95 small-FFT non-AVX load, that's when we're talking!

R23 - Full Load: 51x@1.240v

Light loads: 56x4 - 55x6 - 54x8 - 53x10

I don't think I can maintain full load for hours... But I can play BFV all the night

[RobertoSampaio]

I'm running adaptive=1.524V and V/F#8=1.523V

This pic below if from a 10900K SP63
V/F#8 native = 1.528V

This SP63 running completely Stock will have high Vcore at light loads too.
I think there is no problem with 1.500V at idle.