Power Limit Mod had no effect!

[WebX]

I have a GTX 1070 ROG Strix.

It idles at ~40% power usage, which already seems really high. During testing on something like Kombustor, it spends most of its time at the power limit. It is set to 120% but throttles at ~103%

so using a bit of conductonaut from my 8700K delid, I figured I would give the power limit mod a go.

I coated the two shunt resistors (2mOhm and 5mOhm), booted it up. Absolutely no effect.

So I tried globbing solder to short the resistors entirely. I figure if I create a power fault, I could backtrack from there. Exact same numbers. 40% at idle. throttling easily.

So finally, I completely removed both shunt resistors and bought some 1mOhm resistors to replace them.


EXACT. SAME. NUMBERS. This is not physically possible. Unless all my power is being drawn from the PCI-E slot and I'm not getting power to the molex connector. (That's the only shunt I didn't replace)

Anyone got an idea? I'm clueless.

 

[Johan45]

Software can't really be trusted. I do a fair amount of modding hard and via BIOS. 90% of the time the readouts in GPUz etc.. don't work properly after. The only real way to see if there is any difference is through overclocking. See if you manage to get a bit higher clock speed from it before throttling. I also wouldn't use Kombustor for testing it's just too over the top in power draw to begin with.

 

[WebX]

OK, but in all technicality, I should either be power faulting or seeing 20-50% power usage in Kombustor. Instead, I see that at idle, and my power throttle flag is set right away. Heaven will hold 2012MHz and 1.09V for a few minutes, and then start dropping as per usual to something like 1950 and 1.08V.

 

[UltraTaco]

I raised max allowable voltage in firmware and reflashed the card, but it wouldn't go or allow me to go above defaults. Tried different software, no go. (Gtx460)
Other parameter that I changed worked, so reflash went through, but no voltage increase. Maybe cards have some sort of voltage limiter on lower level that can't be bypassed very easily? Imagine having that on CPUs

 

[EarthDog]

OK, but in all technicality, I should either be power faulting or seeing 20-50% power usage in Kombustor. Instead, I see that at idle, and my power throttle flag is set right away. Heaven will hold 2012MHz and 1.09V for a few minutes, and then start dropping as per usual to something like 1950 and 1.08V.

But is it dropping due to temperature as is normal for these cards? See if the clock drops are close to temperature increases. Just because you are not hitting the power limit doesn't mean something else isn't doing dirty work.

These cards aren't like those in the past... (so thanks Taco, but.......... ).

 

[WebX]

But is it dropping due to temperature as is normal for these cards? See if the clock drops are close to temperature increases. Just because you are not hitting the power limit doesn't mean something else isn't doing dirty work.

These cards aren't like those in the past... (so thanks Taco, but.......... ).

Well that's what I figured given the fact that HWMonitor literally says "TDP Limit" under "Reason for performance limit".

But regardless of what the software is reporting, I physically changed the hardware. The only way the GPU can tell how much power it is drawing is the voltage across the shunt resistors, which would be 1/5th and 1/2 of what they were before.

Also, the temperature is around 71C at most.

 

[Johan45]

Something doen't sound right, even from the start your TDP was very high. Typically with a card in idle you should be at 10% or less. Was this a new card or used?

 

[WebX]

Something doen't sound right, even from the start your TDP was very high. Typically with a card in idle you should be at 10% or less.

Right?! That's more what I'm used to. The only thing I've been messing with are those two shunt resistors.

I bought the card brand new within a week of the card coming out.

 

[Johan45]

Do you have any pics of the work you did? I would also like to see a pic of GPUz sensors tab just to see whats up for myself.

 

[WebX]

Do you have any pics of the work you did? I would also like to see a pic of GPUz sensors tab just to see whats up for myself.

I could pull up the GPUz sensors. But I don't have any pictures of the actual work. I'll have to grab some next time I have the card apart.

I don't really like taking it apart because of wear on the thermal pads, and the fact that despite using lint-free cloth, I get what almost looks like "dust" on the GPU core that doesn't come off with iso or acetone.

 

[WebX]

OK, got the idle power issue figured out I think. It involved enabling "multi-monitor power saving" in Nvidia Inspector. Apparently a problem with 144Hz/G-Sync monitors is that prevents entering idle mode. The core was 1600MHz before, reporting 30-35W of power usage. After using Nvidia Inspector, that dropped to ~8W with a clock of 600MHz. I have yet to test if this affects full 3D performance, but it shouldn't. I totally recommend you guys verify your cards aren't doing this as well.



But anyway, I did find out a bit more. Here are some shots:

EDIT: Taking a second look, I mustve taken these at a weird time, as Heaven was reporting more power usage than Kombustor in these images.


Below is the (new) idle values, which look a ton better.


Here's where things start to get interesting. I had been using afterburner's graphs because I've generally viewed Rivatuner Statistics Server rather favorably.
This is while running Kombustor You can see my power limit flag is being set at only around 100%.


But you can see here, GPU-Z is showing much lower. Usually it was around 60% to 80%. Though the PerfCap reason was still "Power" and was triggered at the same time. I do notice HWMonitor reports the % power usage as "Normalized". I'm not sure what they are normalizing, but that could be the difference between the two.



Finally, I ran heaven again. Clocks have been rock solid and the perfcap reason was "VRel" at 1.09V this time which is exactly what I want.


So I'll run a few more tests, such as a longer Heaven run, and report back if my 'VRel" or "Power" change to "Temp", or if it simply drops my Vcore due to temperature.
I think I'll be fine performance-wise.

I'm just still confused as to why switching from a 5mΩ and 2mΩ shunts to two 1mΩ shunts made basically no difference. It's like Ohm himself is like "naw, my laws don't apply to your 1070". Though perhaps it's a higher-resistance solder. I don't have a sub-ohm meter to find out with.