What NOT to do with your Vega

[MrGenius]

Specifically your Vega FE. But I'm thinking it probably applies to 56 or 64 just the same. Not that there's a lot of members here into hardmodding their Vega 56/64/FE for additional core voltage over stock. But just in case you were/are considering such...or by the off chance that you already have...you should watch this video immediately to avoid potentially expensive GPU core destruction. And sorry @buildzoid, I'm really not trying to rub your nose in it. It was an honest mistake in my book. Could have happened to anyone. I feel for ya. I'm posting this because I genuinely found the vid very informative and wanted to share it here. Lots of good things to know about Vega...besides how to kill it.

 

[buildzoid]

funny thing is I can't replicate the issue. So I still have no idea what caused it.

 

[OneMoar]

craptacular wire insulation combined with the finest chinesium multimeter

it was either a resistance issue or a inductive current issue

and your garbo meter was too slow/inaccurate to catch it

entirely preventable had your meter been good enough you would have caught the vcore issue when it wasn't posting right and NOT killed a 1000 dollar card

 

[buildzoid]

craptacular wire insulation combined with the finest chinesium multimeter

it was either a resistance issue or a inductive current issue

and your garbo meter was too slow/inaccurate to catch it

entirely preventable had your meter been good enough you would have caught the vcore issue when it wasn't posting right and NOT killed a 1000 dollar card

Show me a meter that would pick up on a voltage spike that lasts less than 220 micro seconds and that's including the ramp because from my quick check Fluke doesn't make one.

Also the more annoying thing is I literally can't recreate the issue. So it's not the wire length, trimmer or switch causing the issue. The only way I could get the VRM to shut down on startup was shorting VSEN to GND but that just gave me 1 big spike of 2V before the VRM shut down not 2.

 

[r9]

Show me a meter that would pick up on a voltage spike that lasts less than 220 micro seconds and that's including the ramp because from my quick check Fluke doesn't make one.

Also the more annoying thing is I literally can't recreate the issue. So it's not the wire length, trimmer or switch causing the issue. The only way I could get the VRM to shut down on startup was shorting VSEN to GND but that just gave me 1 big spike of 2V before the VRM shut down not 2.

It's called oscilloscope.

 

[buildzoid]

It's called oscilloscope.

Well yeah and I realized that I should use it far too late.

 

[r9]

Yeah wire length can be a big issue.
I've done quite a few RGH mods on Xbox 360 and there is wire that goes from the cool runner to the reset pad.
You have to play with the wire length to hit the best timing.
Also that slow voltage drop from the second spike to zero talked in the video.
He said that he didn't know what that was.
Well after the second spike the VRAMs turn off and that just the capacitors discharging to 0V.

 

[Aquinus]

Depending on the dielectric, the speed of electricity can vary and going through a DIP switch might alter that enough to delay the signal. I guess that it's possible that during start-up that the frequency of the feedback is having issues with wire and switch you added. I guess there really is no way to know (at least with that card,) now.

 

[buildzoid]

Depending on the dielectric, the speed of electricity can vary and going through a DIP switch might alter that enough to delay the signal. I guess that it's possible that during start-up that the frequency of the feedback is having issues with wire and switch you added. I guess there really is no way to know (at least with that card,) now.

The VRM still fires up even if the core is dead. Also the DIP is there to short VSEN to Vcore(which disables the mod) and the issue was present regardless of the position of the DIP. My most recent theory is that I had an intermittent short on the card that somehow managed to avoid detection by DMM. When I scoped the voltages on the VSEN leg of the pot I didn't see anything wrong. Then again this was after reattaching the mod which somehow fixed the issue so I will never find the cause now.

Oh well from now on when a card isn't getting Vcore after modification I'm not gonna try start it over and over in hopes of having it boot(in this case it would sometimes start up usually after remounting the heatsink).

 

[the54thvoid]

Well, it's people like you that are 'brave' enough to do what you do with expensive hardware that find the things not to do. It's a tech forum and you deserve the TPU Purple Heart for taking one for the team.

And FFS, how old are you? you look about 12! (no offense meant btw).

 

[OneMoar]

he needs a good bench meter
I have the exact meter he uses its a piece of crap my 20 something year old rca is more reliable
a thousand times a good bench meter with frequency and µs ranging will save your bacon

have a 5 thousand dollar inverter motor catch fire is not worth the cost of a 1200 dollar meter

 

[Frick]

he needs a good bench meter
I have the exact meter he uses its a piece of crap my 20 something year old rca is more reliable
a thousand times a good bench meter with frequency and µs ranging will save your bacon

have a 5 thousand dollar inverter motor catch fire is not worth the cost of a 1200 dollar meter

This is getting off topic, but I prefer smaller tools you can move easily. The Fluke 43B is a wonderful thing (apart from the chargers frying out like nothing I've ever seen).

 

[OneMoar]

build is right there isn't a hand held meter in existence that will read down to 220µs the flukes go down to about 250µs in fast peak mode
a good bench meter should read down to 100µs if it has frequency capability
we use keysights and flukes at work for our portables

 

[eidairaman1]

Do you think a bigger gauge of wire might fix the resistance delays?

 

[OneMoar]

Do you think a bigger gauge of wire might fix the resistance delays?

you want stranded wire the smallest gauge you can work with with good insulation

larger wire = more resistance,at least in this application

 

[buildzoid]

Well, it's people like you that are 'brave' enough to do what you do with expensive hardware that find the things not to do. It's a tech forum and you deserve the TPU Purple Heart for taking one for the team.

And FFS, how old are you? you look about 12! (no offense meant btw).

I can never miss a chance to use this:

I recommend not searching for the source.


My most recent theory on what killed the card is that I had some kind of short on the card that only occured when the card was mounted in the system or something. Because shorting Vsense to GND is the only way I can get the VRM to shut down again but I still can't replicate the 2 voltage spikes that look like some kind of ringing so I have no idea what's up with that.

As for a better meter...

I can't really afford one and I realize how that sounds after killing a 1000GBP GPU but I could've saved the card with scope which I already have. On the other hand I would really like to get a multimeter that does milliohms. A vega GPU core has a resistance of 0.2ohms according to the DMM I have now and on the Nvidia side GP104 is around 90milliohms and GP102 is even less.

 

[OneMoar]

could just put it on frequency and do the math

 

[buildzoid]

could just put it on frequency and do the math

Except I don't have a frequency to measure. The Vsen line under normal condition should literally be equal to Vcore or a fraction of Vcore. It's not some high frequency signaling line(the IR 35217 only checks the line at 20.8KHz). The only time it should exhibit any delay problems is when Vcore is changing like at start up or on power state transitions.

 

[Mordaur]

funny thing is I can't replicate the issue. So I still have no idea what caused it.

Sorry to bring this back from the dead, but i have a spare vega 64 that i wanted to voltmod, and there is not much to be found on the matter, besides your work.
the interesting thing is the reply from Haruhi Shimasaki on above mentioned YT video, and I quote:

Wait, did the control loop's resistance go up significantly (like 10~100x as high)?
If so, I think that the problem was the feedback loop's resistance and, I assume, a smoothing capacitor on the voltage controller's feed back loop made a low pass filter. The R and C values were (probably) chosen so that the ~300kHz switching frequency would be filtered out to get a better average voltage reading. At this level of filtering, there should not be a significant delay.

HOWEVER, if the R value is increased by a factor of 100, transient delays will also increase by 100.

So assuming the 2 resistors you removed from the sense circuit do indeed act as a LPF/RC (resistor-capacitor) Integrator Circuit, and you skew the series resistance before the capacitor, you are skewing the voltage read by the sense pin, but are also skewing the transient delay/ramp up time of the filtering circuit. which does explain the over/undershoot.

ideally it would be better to cut the positive trace going from the capacitor to the sense pin, and put the potmeter in-between that. you get the clean filtered vcore voltage, and skew that after the capacitor. Or one could potentially remove the capacitor (and lose out on smoothing/filtering), but that would resolve the filtering transient delay of the circuit.

i'm not an engineer, but it does make sense like this, and could potentially save a card here and there if someone should still follow your voltmod tutorial (besides using a EVC2).
it would have been nice to see the oscilloscope reading before and after the potmeter at the capacitor. This could visually represent and compare the ramp up of what the vrm actually outputs, and what the sense pin reads.

I have a asus strix vega so i have the luxury of the hotwire header, i'll have to see what can be done with that. none ever documented the header, and how its electrically connected.
The pcb also seems to have a I2C header near the controller and I have a EVC2 incomming.

Just thought i'dd share.

 

[eidairaman1]

Go to Igorslab and look there for mods

 

[Mordaur]

Go to Igorslab and look there for mods

unfortunately i cant find anything there. Searching for vega+voltmod on their forum, or through google site search yields nothing, or nothing interesting. Most post are about undervolting. i don't want to undervolt. my vega is perfectly stable at 1725 core (boostst to 1790), and 1070 hbm2 at 1200mv. if I try anything above 1725 core (lets say 1735), it'll try to boost past 1800 and crash, so the core is hungry for more voltage. i'm sure i can push it more, if only I could give it more voltage.
This is a strix vega 56 actually, with 64 bios.

 

[TheoneandonlyMrK]

unfortunately i cant find anything there. Searching for vega+voltmod on their forum, or through google site search yields nothing, or nothing interesting. Most post are about undervolting. i don't want to undervolt. my vega is perfectly stable at 1725 core (boostst to 1790), and 1070 hbm2 at 1200mv. if I try anything above 1725 core (lets say 1735), it'll try to boost past 1800 and crash, so the core is hungry for more voltage. i'm sure i can push it more, if only I could give it more voltage.
This is a strix vega 56 actually, with 64 bios.

Through much messing about with a frontier edition 8GB bios with an extra 50mV on its scale I can tell you not much they're at their limits already but I did see 1803 stable, beyond that hard failed.

 

[eidairaman1]

unfortunately i cant find anything there. Searching for vega+voltmod on their forum, or through google site search yields nothing, or nothing interesting. Most post are about undervolting. i don't want to undervolt. my vega is perfectly stable at 1725 core (boostst to 1790), and 1070 hbm2 at 1200mv. if I try anything above 1725 core (lets say 1735), it'll try to boost past 1800 and crash, so the core is hungry for more voltage. i'm sure i can push it more, if only I could give it more voltage.
This is a strix vega 56 actually, with 64 bios.

That's normal

 

[damric]

I never bothered with going above the stock limit of 1.2v on either my Vega56 or 64, but I did mess around with hex editing registry values for current and power to get them well over 400W for some under water benching fun. If I recall, the Vega 64 of mine could sustain about 1800MHz for some of the 3DMarks, but pulling crazy amount of power. I could only imagine if I would have put one of the BIOSes that have 1.25v. If you want my modified power tables, I'll link them. I took Buildzoids and pushed them much higher. I was seriously worried about burning the PCIE power connectors when benching that high on those cards.

 

[Mordaur]

I never bothered with going above the stock limit of 1.2v on either my Vega56 or 64, but I did mess around with hex editing registry values for current and power to get them well over 400W for some under water benching fun. If I recall, the Vega 64 of mine could sustain about 1800MHz for some of the 3DMarks, but pulling crazy amount of power. I could only imagine if I would have put one of the BIOSes that have 1.25v. If you want my modified power tables, I'll link them. I took Buildzoids and pushed them much higher. I was seriously worried about burning the PCIE power connectors when benching that high on those cards.

i've made my own with a 250% power target (extracted from the bios with OverdriveNTool, modified with Vega64SoftPowerTableEditor). but it pulls max 330w (according to gpuz/hwinfo). superposition score seems low, but could be due to windows 11
Since its originally a asus strix vega 56, i've only the 64 (260w) bios available. unfortunately no LC bios.
Not sure if the current limit scales with the power offset. hwi seems to believe so, but the card won't pull anything more.