Overclocking CPU Cache

[Cmptr]

CPU: 11900K

MOBO: Asus Maximus HERO XIII

 

Vcore mode: adaptive with +0.001 offset

XMP 1

5.2Ghz all Core

AVX2 and AVX512 Offset: 0

AC/DC: 0.001 (Or 0.1 or whatever the lowest)

LLC: 8

VID for 5.1: -0.15 (Only to allow VID for 5.2 to be set)

VID for 5.2: -0.22

 

Average vcore under load: ~1.394

Average vcore idle: ~1.403

Max vcore: 1.421

 

Max CPU package power: 413

Average: 374

 

Before I enabled AVX, uncore was 4.3

After enabling it was 4.0

 

Tests I ran:

Without AVX:

Blender BMW for 20 hours (That was however with LLC4 and VID for 5.2 -0.04.  Only difference for vcore is that idle was 1.51 and instant 413W within less than 1 min full load.  Whereas with settings above, I only saw 413 with AVX both enabled)

 

Last night I ran ffmpeg encoding a 4K movie using libx265 at preset slow.  8 hours with settings above and no issues.  

 

Now onto uncore..

 

There is this setting I needed to disable in bios to actually be able to get anything higher than 4.0.  

 

When doing so an setting uncore to 4.3, obviously vcore goes up.  This case it went to ~1.43.  Didn't test for long.  

 

There are these other voltages in bios such as System Agent, CPU VCCIO, Mem VCCIO etc.  (There was also one that had the option of using adaptive)

 

I know most of them are for RAM, (Atleast that is what I am reading, but I am reading that some of that has changed)

 

What I'm most curious about is, is there a voltage that I can change that will "Only" (I and I use "Only" very loosly) that will help cpu cache and not raise the vcore?  Or no matter what I change, vcore will be affected?

 

Now I know there isn't much benefit at all for higher CPU cache for most situations, but I still want to OC it.

 

Any help is much appreciated.  (Sorry if this post is all over the place)

[RONOTHAN##]

9 minutes ago, Cmptr said:

Now I know there isn't much benefit at all for higher CPU cache for most situations, but I still want to OC it.

It helps in a lot of multi core workloads and anything where the memory helps a lot, it is actually a good idea to OC it, it's just not exactly fun to stress test.

 

VCCSA and VCCIO do power the cache somewhat so increasing those can help stabilize it, but VCore affects it the most. 

 

I would recommend you stick to a fixed VCore, the voltage stability it brings helps a fair bit with system stability and the power draw at idle concerns don't really matter given the values you're already seeing at idle. It might get you an extra 100-200MHz on UCLK for free. 

[Cmptr]

8 minutes ago, RONOTHAN## said:

It helps in a lot of multi core workloads and anything where the memory helps a lot, it is actually a good idea to OC it, it's just not exactly fun to stress test.

I mainly just do coding, (Usually C++) and some video encoding focusing on compression using HEVC.  Sometimes I do video editing and 3D designing. 

10 minutes ago, RONOTHAN## said:

 

I would recommend you stick to a fixed VCore, the voltage stability it brings helps a fair bit with system stability and the power draw at idle concerns don't really matter given the values you're already seeing at idle. It might get you an extra 100-200MHz on UCLK for free. 

The voltages I'm seeing at idle are only that high because I have c-states disabled and have min CPU state in power plan in windows at 100%.  These settings are only temporary.  Once I finish overclocking, I'm going to enable c-states and set min CPU state to 0%.

 

Doing that, idle vcore will be under 1v.  That is why I don't want to use fixed vcore.  

14 minutes ago, RONOTHAN## said:

VCCSA and VCCIO do power the cache somewhat so increasing those can help stabilize it, but VCore affects it the most. 

How can I lower what the vcore will be?  I tried lowering VID for 5.2 even more, and that had no effect on what vcore was.

[RONOTHAN##]

1 minute ago, Cmptr said:

I mainly just do coding, (Usually C++) and some video encoding focusing on compression using HEVC.  Sometimes I do video editing and 3D designing. 

17 minutes ago, RONOTHAN## said:

Code compilation tends to really benefit from fast cache IIRC, so yeah it's worth at least taking a look at it. 

 

3 minutes ago, Cmptr said:

The voltages I'm seeing at idle are only that high because I have c-states disabled and have min CPU state in power plan in windows at 100%.  These settings are only temporary.  Once I finish overclocking, I'm going to enable c-states and set min CPU state to 0%.

 

Doing that, idle vcore will be under 1v.  That is why I don't want to use fixed vcore.  

I get that, but at the same time adjusting VID settings introduces more instability and is a lot more of a headache. The 2-3W lower idle consumption IMO isn't worth the headache and lower stability of using an offset rather than a fixed offset. 

[Cmptr]

6 minutes ago, RONOTHAN## said:

I get that, but at the same time adjusting VID settings introduces more instability and is a lot more of a headache. The 2-3W lower idle consumption IMO isn't worth the headache and lower stability of using an offset rather than a fixed offset. 

Right now at idle, I think I'm at ~137W, you're saying even with c-states and a different power plan in windows, it will only be 2 - 3 watts lower?  I honestly thought it would be lower.  I'll have to test that.

[RONOTHAN##]

1 minute ago, Cmptr said:

Right now at idle, I think I'm at ~137W, you're saying even with c-states and a different power plan in windows, it will only be 2 - 3 watts lower?  I honestly thought it would be lower.  I'll have to test that.

137W at idle CPU power is very high considering. It should be closer to 50W-60W. Test it, but that number doesn't really sound right. 

[Cmptr]

2 hours ago, RONOTHAN## said:

137W at idle CPU power is very high considering. It should be closer to 50W-60W. Test it, but that number doesn't really sound right. 

At idle right now without c-states and min cpu state at 100%, the min CPU Power Package is at is 112.  Changing min cpu state to 0%, I get min 86.  With C-States set to auto, I get min 56..  However, idle voltage is 1.43, and I'm guessing this is because of LLC being at 8?..

[RONOTHAN##]

6 minutes ago, Cmptr said:

However, idle voltage is 1.43, and I'm guessing this is because of LLC being at 8?..

Sounds about right. That's not unsafe or anything, 1.43V is perfectly acceptable for idle voltage, I'd be fine with anything upwards of 1.47V idle as long as under load it wasn't thermal throttling. Personally I'd aim for the droopiest LLC that you can have while still maintaining a somewhat reasonable idle voltage, which it sounds like you've found. 

[Cmptr]

3 minutes ago, RONOTHAN## said:

Sounds about right. That's not unsafe or anything, 1.43V is perfectly acceptable for idle voltage, I'd be fine with anything upwards of 1.47V idle as long as under load it wasn't thermal throttling. Personally I'd aim for the droopiest LLC that you can have while still maintaining a somewhat reasonable idle voltage, which it sounds like you've found. 

Changing to LLC6, I'm seeing idle voltages under 1v now though.  To be specific, it's about 0.746...  Seems like now I'm kind of back to the drawing board.

 

Isn't 1.4 idle at 800MHz bad?

[RONOTHAN##]

Just now, Cmptr said:

Changing to LLC6, I'm seeing idle voltages under 1v now though.  To be specific, it's about 0.746...  Seems like now I'm kind of back to the drawing board.

 

Isn't 1.4 idle at 800MHz bad?

It's higher than necessary, but I've run chips like that with disabled C states on more voltage than that at idle for pretty long. As long as it's not super high (>= 1.48V idle) I wouldn't worry about it. 

 

It sounds like voltage offsets were just broken in LLC level 8 though. Again, I'd stick to a static voltage though, it decreases transients and improves stability a fair bit. 

[Cmptr]

32 minutes ago, RONOTHAN## said:

It's higher than necessary, but I've run chips like that with disabled C states on more voltage than that at idle for pretty long. As long as it's not super high (>= 1.48V idle) I wouldn't worry about it. 

 

It sounds like voltage offsets were just broken in LLC level 8 though. Again, I'd stick to a static voltage though, it decreases transients and improves stability a fair bit. 

Well LLC is known for keeping a high Idle Voltage.  I have yet to try LLC7 to see if I get the same as I did with LLC6.  But right now, I'm using FFmpeg with uncore at 46 to see what happens.  Though that may be pointless as I'm not sure if FFmpeg benefits from fast memory..

[RONOTHAN##]

39 minutes ago, Cmptr said:

Well LLC is known for keeping a high Idle Voltage.  I have yet to try LLC7 to see if I get the same as I did with LLC6.  But right now, I'm using FFmpeg with uncore at 46 to see what happens.  Though that may be pointless as I'm not sure if FFmpeg benefits from fast memory..

I know it keep high idle voltage, it's just that if you're using a voltage offset like you said you were, it should still be around 1V instead of 1.4V. It sounds like it was set at a static voltage of 1.45V (LLC level 8 should have VDroop, even at idle)

 

I don't really use FFMpeg for anything, so can't comment on how much Uncore helps with that. Benchmarks like Y Cruncher, SuperPi, 3DMark CPU benchmarks, and a number of others do benefit a lot from high uncore. 

[Cmptr]

14 minutes ago, RONOTHAN## said:

I know it keep high idle voltage, it's just that if you're using a voltage offset like you said you were, it should still be around 1V instead of 1.4V. It sounds like it was set at a static voltage of 1.45V (LLC level 8 should have VDroop, even at idle)

 

I don't really use FFMpeg for anything, so can't comment on how much Uncore helps with that. Benchmarks like Y Cruncher, SuperPi, 3DMark CPU benchmarks, and a number of others do benefit a lot from high uncore. 

LLC8 should have no vdroop as it is the highest level on my board.  There should just be transients  I think is what they are called.  That or is it called just dips and spikes?

 

Thank you for those benchmarks, I will for sure try them.

 

Also, I did not have any static voltage set whatsoever.  It has always been on adaptive.

[RONOTHAN##]

7 minutes ago, Cmptr said:

LLC8 should have no vdroop as it is the highest level on my board.  There should just be transients  I think is what they are called.  That or is it called just dips and spikes?

 

Thank you for those benchmarks, I will for sure try them.

 

Also, I did not have any static voltage set whatsoever.  It has always been on adaptive.

I forgot ASUS does LLC backwards. I'm more used to MSI and ASRock boards with Level 1 being no VDroop and Level 9 being no LLC. You want to avoid Level 8 (Asus), it's the noisiest option, and will require much higher VCore to stablize. Level 2 or 3 is likely the most optimal. 

[Cmptr]

3 hours ago, RONOTHAN## said:

I forgot ASUS does LLC backwards. I'm more used to MSI and ASRock boards with Level 1 being no VDroop and Level 9 being no LLC. You want to avoid Level 8 (Asus), it's the noisiest option, and will require much higher VCore to stablize. Level 2 or 3 is likely the most optimal. 

That is not accurate.  In ASUS Bios, as you get lower, you get more vdroop.  If that wasn't the case, then LLC8 would not yield me pretty much the same vcore for load as idle. 

[RONOTHAN##]

24 minutes ago, Cmptr said:

That is not accurate.  In ASUS Bios, as you get lower, you get more vdroop.  If that wasn't the case, then LLC8 would not yield me pretty much the same vcore for load as idle. 

That's what I said, level 8 has no VDroop, the idle and load voltages are the same. The problem with LLC effectively maxed out like it is on Level 8 is that you end up with a very noisy VCore rail. 

 

This is on a X570 Godlike (IIRC) instead of the Maximus XIII Hero, but the same principles apply. Also MSI does VDroop backwards, Level 1 is flat, level 9 is lots of VDroop.