9900k tweaking

[splerdu]

So both Auto and Normal VCore in BIOS meant the CPU was asking for way too many volts. VID would be as high as 1.45V+. Did a manual OC by setting 5GHz and 1.35V, which led to a measured 1.28V under load (VR VOUT from ISL69138). Temps are now much more controlled and the CPU can sustain 5GHz throughout a Cinebench run giving much improved performance.



Previously my 5GHz OC via MCE would result in a score varying anywhere between 4600 and 4880 because the high volts meant it was completely limited by thermals. With the manual OC temps are just a hair below 100C and the CPU manages to avoid thermal throttling.

Running stability test now with Prime95, then it's off to find the right offset voltage to get variable VID while hopefully keeping temps under control same as with this config.

[splerdu]

Playing around with offset voltages, since manual OC means the CPU will be running 1.28-1.35V all the time.

Since manual OC showed that asking for 1.35V was stable, theoretically I could go back to MCE on, VCore Normal, and VID offset -0.100 (1.45V VID - 1.35V BIOS VCore for manual OC). Problem would be when the core downclocks to 800MHz and is asking for 0.700 volts. Not sure 0.600V will be stable.

Challenge then is to find a VID offset that brings VCore low enough under load that the CPU doesn't throttle too much, but high enough under no-load so it stays stable even when in a low-power state.

-0.050V seems to be stable under Prime95 and IBT, but it is seeing throttling under sustained Prime95 (eventually the 240mm rad isn't enough for the sustained heat output). Peak voltage during overshoot is 1.411V, some 0.025V higher than with the manual OC.



Going to try -0.060V, and -0.070V later...

Measured VIDs for reference:

[splerdu]

Memory tweak time!

TPU published a Ryzen memory overclocking guide where the SK Hynix equipped 3600CL19 G-Skill have been shown to be excellent overclockers. Followed the timings from Ryzen DRAM calculator and got it running at 3600CL16-19-19-36:

[splerdu]

More memory tuning: Tried 3700, 3733 and 3800.

TestMem picked up 1 error at 3800. Seems like could be easily fixed with a slight voltage bump, but I'll probably leave this at 3733 instead.

VID offset testing continued: -0.055 stable. Trying -0.060. Finding it a bit funny that stability under idle is the concern, rather than stability under load.

[borgy_t]

how's the real-world gains?

[splerdu]

The memory OC? Parang not much lol. At least nothing I can feel with my 1440P 60Hz monitor 

The VID offset is much more useful. Good cores now top out at 90C in Linpack, only two cores reach 99C, others peak at 97C. Big boost for workloads like Handbrake because the CPU sustains higher clocks all throughout.

Average voltage is now less than 1.2V, peak voltage (including overshoot) is 1.38V.

[borgy_t]

ah, so decent improvement in rendering times, next na ba solder TIM?

[splerdu]

You mean delid and solder the processor? The 9000 series processors are already soldered.

It is possible to delid and lap the die itself but that's rather hardcore already. Going by logical steps full custom loop comes before even thinking about doing that.

[borgy_t]

oops i meant delid it and replace the TIM with liquid metal hehehe

[splerdu]

Yeah it's soldered already (again) so only minor gains to be made doing that. GamersNexus found a little less than 4C improvement going from solder to LM, and 5C worse going from solder to high-end paste.



Considering the risk of cracking the die while doing the delid it's more logical to upgrade the water loop first, maybe a 280mm or 360mm rad.

https://www.gamersnexus.net/guides/3381-intel-i9-9900k-delid-liquid-metal-5-4ghz-overclock