[Timur Born]

Tried it and didn't work. Power draw 2 much and when locked at 250W the cinebench score about 37K.

So far best performance with AC/DC 0.1/1 m0hm.

Increasing DC is the only factor I noticed that reduces power draw in my cpu.

Is there a con to using a DC of 1 m0hm

Maybe your CPU LLC was too high then or your board's stock DC is too far out there. I get 40-41k CB23 scores at 235-245W (still working on that).

 

[SafetyHammer]

Maybe your CPU LLC was too high then or your board's stock DC is too far out there. I get 40-41k CB23 scores at 235-245W (still working on that).

What's the AC DC are you using?
Did you set current limit or not?

 

[Timur Born]

No current limit, because it didn't do what I expected it to and power limit is good enough for me.

Originally I started with a BIOS AC/DC preset "Power Saving", which sets 30/90 (0.300/0.900 mOhm) in combination with a -0.08v offset. That already pushes CB23 power from 290W to 240-245W. Currently I am experimenting with AC values between 1-10 and correspondingly lesser offsets. This effectively increases single/low core Vcore while keeping all-core load about the same.

CB23 is stable down to about 230W for me, but getting all the other load situations stable alongside is the tricky part, especially in combination with per-core OC. Currently I am at about 236W for CB23 all-core and 38-40W for single-core at 6 GHz (1.35v) below 65°C. Folding@home stability around the TVB downbin point of 65°C is a problem, though, so I am still working on that one.

 

[SafetyHammer]

Folding@home stability around the TVB downbin point of 65°C is a problem, though, so I am still working on that one.

This part I don't understand.

Also did you keep dc as 90?

 

[Timur Born]

Yes, I kept my DC at stock 90, only working with AC, offset and CPU LLC here, which in combination with power limit and TVB are more than enough adjusting screws to get confused by.

 

[Timur Born]

You can ignore the TVB part at stock frequencies. I am experimenting with higher per core OCs and use Thermal Velocity Boost to downclock above a temperature limit (to get back to stock clocks).

 

[SafetyHammer]

No current limit, because it didn't do what I expected it to and power limit is good enough for me.

Originally I started with a BIOS AC/DC preset "Power Saving", which sets 30/90 (0.300/0.900 mOhm) in combination with a -0.08v offset. That already pushes CB23 power from 290W to 240-245W. Currently I am experimenting with AC values between 1-10 and correspondingly lesser offsets. This effectively increases single/low core Vcore while keeping all-core load about the same.

CB23 is stable down to about 230W for me, but getting all the other load situations stable alongside is the tricky part, especially in combination with per-core OC. Currently I am at about 236W for CB23 all-core and 38-40W for single-core at 6 GHz (1.35v) below 65°C. Folding@home stability around the TVB downbin point of 65°C is a problem, though, so I am still working on that one.

Managed to get it to 243W however P core speed drop to 5.4 instead of 5.5 and E drops to 4.2 instead of 4.3Ghz while running Cinebench R23, Score 39.5K. Couldnt figure out how to keep it at 5.5/4.3Ghz at that power.

 

[SafetyHammer]

40 is the impedance value for this particular board's LLC6 setting. DC LL is used to calibrate, or match the VIDs to the Vcore under full load. Here are some other DC LL Values:

LLC5= DC LL 10
LLC6= DC LL 40
LLC7= DC LL 69
LLC8= DC LL 98

Trying to find the source of the LLC/DC values you mentioned for MSI, I can only find for Asus.

Would it be possible to share the source please

 

[Titan2t]

Thank you for the great discussion! Can someone give me some guideance, please?

I have a 13900k on a Gigabyte Aero G Z790 and would like to limit its max temp to about 80 degrees but at the same time lose as little performance as possible. How do I approach that?

If I simply set PL1 and PL2 a limit that works well for the temperature. But is there a way to increase performance at the same time?

 

[persizi]

Increasing DC is the only factor I noticed that reduces power draw in my cpu.

The power draw is the same, just the reading are incorrect. For the most correct power reading AC and DC Lite Load must have the same values.

 

[walliart]

The power draw is the same, just the reading are incorrect. For the most correct power reading AC and DC Lite Load must have the same values.

I am also trying to undervolt a 13900KF, just to keep the temperatures lower. I put a contact frame and thermal grizzly kryonaut extreme (instead NH1 stock paste that comes with Noctua NH D15) and this gave me about 10C lower temperatures.
System is like this: i9-13900KF stock; Gigabyte (BIOS F5m) Aorus Master Z790; Kingston FURY Renegade Silver 32GB DDR5 6400MHz CL32; NH-D15 Chromax.Black
What I have modify in BIOS:
Voltage 1.335V adaptive.
Negative offest:
-0.115V at VF point 9 (58x)
-0.100V at VF point 8 (57x)
-0.090V at VF point 7 (54x)
-0.050V at VF point 6 (51x)
CPU internal AC/DC Load line on PowerSaving
CPU Vcore Loadline Calibration LLC is on Turbo
AC LL 8
DC LL 100
Disable for undervolt protection and at voltage reduction initiated TVB
Enhanced Multi core performance auto

For this combination I have obtained the lowest power consumption ~ 250W, but temperatures are around 95C in 10 min Cinebench R23

I don't know why the readings are for Core Vid 1.055V and Vcore 1.272V

If I go with AC LL 1 and DC LL 1 like in "SkatterBencher #49"
I get CoreVid 1.253V and Vcore 1.248V that are more alike, temperatures are slightly lower, under 90C, but the power is 285W. How can higher power give me lower temperature I don't understand.

Now, which of these two settings are better, or how do I improve them please?

 

[persizi]

If I go with AC LL 1 and DC LL 1 like in "SkatterBencher #49"
I get CoreVid 1.253V and Vcore 1.248V that are more alike, temperatures are slightly lower, under 90C, but the power is 285W. How can higher power give me lower temperature I don't understand.

When the AC and DC Lite Load values are not the same, power readings are inaccurate.
I would suggest better stick with AC/DC LL 1 and adjust the voltage from the LLC.

 

[xh43k]

When the AC and DC Lite Load values are not the same, power readings are inaccurate.
I would suggest better stick with AC/DC LL 1 and adjust the voltage from the LLC.

Seems like bullshit, because mobos put like 50/80 as default AC/DC
So I guess they wouldn't put wrong auto values to cause wrong voltage readings.

1/1 value will definitely cause wrong voltage readings though.

 

[persizi]

The voltage readings are fine. The power readings are incorrect. Everyone can check this with a wattmeter on the wall. 50/80 are not very different like 8/100, so the deviation is not big.

 

[SafetyHammer]

Seems like bullshit, because mobos put like 50/80 as default AC/DC
So I guess they wouldn't put wrong auto values to cause wrong voltage readings.

1/1 value will definitely cause wrong voltage readings though.

Fyi
Mode 1 is 1/1,
Mode 2 is 10/80,
Mode 3 is 15/80
Mode 4 is 20/80
Mode 5 is 25/80
Mode 6 is 30/80
Mode 7 is 35/80
Mode 10 is 60/80
Mode 12 is 80/80
Mode 16 is 120/120

for Z790 TOMAHAWK DDR4

 

[xh43k]

Fyi
Mode 1 is 1/1,
Mode 2 is 1/80,
Mode 3 is 15/80

for Z790 TOMAHAWK DDR4

Yeah I know because I have ddr5 version of tomahawk, except mode 2 is 10/80 here, then each node adds 5 to AC so:

1 1/1
2 10/80
3 15/80
4 20/80
etc

Thus why mode 1 provides incorrect readings

 

[SafetyHammer]

Yeah I know because I have ddr5 version of tomahawk, except mode 2 is 10/80 here, then each node adds 5 to AC so:

1 1/1
2 10/80
3 15/80
4 20/80
etc

Thus why mode 1 provides incorrect readings

It was a typo, mode 2 also 10/80.

I use mode 2 or advanced 10/100 which gives the best performance for me in LL6 with adaptive plus minus offset 0.06

 

[persizi]

It was a typo, mode 2 also 10/80.

I use mode 2 or advanced 10/100 which gives the best performance for me in LL6 with adaptive plus minus offset 0.06

I would suggest Mode 1, LLC5 and if needed adjust the voltage through the offset method not the adaptive.
Mode 1 will give you higher power reading in HWiNFO compared with Mode 2 but this is not accurate. Actually the power is less because of the lower voltage.
Or if you are worried about the power reading you can keep DC Loadline to 80 or even put it to 110, that will give you even lower power readings

 

[SafetyHammer]

I would suggest Mode 1, LLC5 and if needed adjust the voltage through the offset method not the adaptive.
Mode 1 will give you higher power reading in HWiNFO compared with Mode 2 but this is not accurate. Actually the power is less because of the lower voltage.
Or if you are worried about the power reading you can keep DC Loadline to 80 or even put it to 110, that will give you even lower power readings

At the moment with mode 2 or 10/100 with LLC 6, Adaptive plus offset - 0.06 my voltage sits on 0.8/1.28 idle/gaming and drop to 1.23 in cinebench.

This is with P 5.5, E4.3 and Ring at 4.5. If I have 2 cores goes to 5.8 then voltage is about 1.3

Power limit 253, Current limit 400

 

[persizi]

Adaptive plus offset - 0.06 my voltage sits on 1.29/1.28 idle/gaming

On idle the voltage should go below 0.800, that's why I suggest using regular offset instead Adaptive. But it is more complex to keep the rest of the voltage in the same order.