Question about System Agent Voltage

[Vynra]

specs: 13900k, rog strix E gaming Wifi, 7600mhz gskill ram

so i decided to tinker and lowered system agent voltage from what looked like close to 1.4 (this is what auto was pushing) down to fixed voltage of 1.25. i left everything else auto including imc and vddq since im using "xmp tweaked" option.

my question is everything seems stable(tested with TM5 absolut and kahru) but not only that im getting better benchmarks in timespy  tests and what seems like faster/smoother feeling during games. the only downside i saw is i saw cpu VID(not vcore. actual vcore was only 1.41) hit over 1.46v. the most i saw it hit before the change was a little over 1.4ish on the vid.never seen it above 1.45. is this common for lowering system agent voltage?

any insight would be appreciated

[Silent_Scone]

Hi @Vynra 

Changing VCCSA alone should have no impact on the requested VID. The amount of VCCSA required for stability at a given frequency is largely dependant on the CPU and to a lesser extent the memory kit and motherboard. Auto values often account for worst case scenario in respect to the quality or rather capability of the CPU silicon.

[Vynra]

which i do find odd which is why i brought it up but what about better benchmakrs/smoother gameplay? is that normal?

[Silent_Scone]

How did you test system stability? Could be a case of error correction resulting in worse 1% lows.

[Vynra]

i tested with occt cpu test avx2, tm5 absolut, kahru.. also benchmarked with timespy 3d mark

[Silent_Scone]

Any rail carrying a signal like this (VCCSA) can be finicky, hence why applying certain voltage values can hurt stability.

[Vynra]

So do you think this is too low of a voltage or am I actually getting this performance uplift? I went from a 24000 in timespy cou test to a 26000. I’ve never seen that number before. 

[Silent_Scone]

If the system is stable then it’s not too low a value, no. The requirement for a given frequency is individual to that CPU.

[Silent_Scone]

The 1.34v profile is not entirely stable, however, it does have an impact on the score (Timespy Havok based physics tests very latency sensitive).

The requested VID for the CPU stays the same, as expected.

VCCSA 1.34v

 VCCSA [Auto] 1.23v

 

[Vynra]

wow thanks so much for testing it out! i appreciate it. so let me ask you how do you know that the 1.34 profile is not entirely stable? it looks like you got a performance uplift.

and you are correct, i switched cpu system agent voltage back to auto and cpu VID didnt change. i also did timespy again and the score didnt change. i think either a 3d mark update or windows update made the score go up.