GTX 1080 volt issues past 1.25v

[othertomperson]

1 hour ago, boboman342 said:

PC Per's interview with an Nvidia dude said clock for clock (shader per shader), pascal is the same performance as maxwell.

Overclock a 980 to 2GHz and see how well that goes. Dramatic clock speed increases are part of the improvements of Pascal over Maxwell. You can't just artificially limit them for Pascal while allowing Maxwell to run at stock and then claim that they are therefore the same.

[done12many2]

3 hours ago, Lays said:

Wouldn't be possible with this volt issue.

 

If people are Ln2'ing cards and only hitting 2300-2400 on lots of cards tested, than water won't get anywhere near those clocks without a fix for this volt issue.

 

Have you noticed a trend with 95% + ASICs with Pascal.  I've only seen 3 ASICs so far and all of them were 95% +.  It's my understanding that higher scoring chips don't generally respond to voltage as well as lower scoring chips, but on average require less voltage at a given frequency.  With this said, is it possible that there is a higher quality being pushed, which may have resulted in nvidia pushing this 1.25v lock?

 

This could also explain the poor scaling at higher clock rates/voltage.

[Stefan1024]

When lowering the production process pitch, the voltage required to switch the FET goes down.

Therefor if you have the same voltage on maxwell and pascal, the pascal GPU is actually much more overvolted.

[Stefan1024]

1 minute ago, done12many2 said:

 

Have you noticed a trend with 95% + ASICs with Pascal.  I've only seen 3 ASICs so far and all of them were 95% +.  It's my understanding that higher scoring chips don't generally respond to voltage as well as lower scoring chips, but on average require less voltage at a given frequency.  With this said, is it possible that there is a higher quality being pushed, which may have resulted in nvidia pushing this 1.25v lock?

 

This could also explain the poor scaling at higher clock rates/voltage.

95% on a new process for a middle sized chip? Not happeing if they didn't set the threshold for "100%" to a rather low value.

[done12many2]

3 hours ago, Stefan1024 said:

95% on a new process for a middle sized chip? Not happeing if they didn't set the threshold for "100%" to a rather low value.

 

Umm...  Okay?

 

 

[Lays]

4 minutes ago, Stefan1024 said:

When lowering the production process pitch, the voltage required to switch the FET goes down.

Therefor if you have the same voltage on maxwell and pascal, the pascal GPU is actually much more overvolted.

Does the same not apply to haswell and skylake then?  Skylake has higher voltage tolerance, and higher VID than Haswell chips did.

[DXMember]

22 minutes ago, othertomperson said:

Overclock a 980 to 2GHz and see how well that goes. Dramatic clock speed increases are part of the improvements of Pascal over Maxwell. You can't just artificially limit them for Pascal while allowing Maxwell to run at stock and then claim that they are therefore the same.

but.. but.. they are the same... it's just a die shrink it's all TSMC FF

where's the 5 miracles of Pascal we were promised? Where's Pascal = 10x Maxwell? Where's HBM2 for god's sake?

[DXMember]

10 minutes ago, Lays said:

Does the same not apply to haswell and skylake then?  Skylake has higher voltage tolerance, and higher VID than Haswell chips did.

massively depends on the architecture, but if they were the same then yes - it would apply

AMDs/TSMCs 32nm chips take a whole lot more voltage than Intels Sandy Bridge does

[Stefan1024]

9 minutes ago, done12many2 said:

 

Umm...  Okay?

 

The ASIC quality get measure after production and the guideline for this test is nowhere published. Nvidia can copute the percentage however they like. So 100% is probably not a "perfect" chip with no defect what so ever.

Or the reviewer cards are just cherry picked.

 

8 minutes ago, Lays said:

Does the same not apply to haswell and skylake then?  Skylake has higher voltage tolerance, and higher VID than Haswell chips did.

The voltage / size coupling is given by physics. The difference can occure from the change from integrated VRMs to the external. Buffer for losts in the traces.

[othertomperson]

11 minutes ago, DXMember said:

but.. but.. they are the same... it's just a die shrink it's all TSMC FF

where's the 5 miracles of Pascal we were promised? Where's Pascal = 10x Maxwell? Where's HBM2 for god's sake?

Nvidia never claimed that the 1080 was based on the GP100 chip, which is what that marketing event was all about.

[done12many2]

3 hours ago, Stefan1024 said:

The ASIC quality get measure after production and the guideline for this test is nowhere published. Nvidia can copute the percentage however they like. So 100% is probably not a "perfect" chip with no defect what so ever.

Or the reviewer cards are just cherry picked.

 

The voltage / size coupling is given by physics. The difference can occure from the change from integrated VRMs to the external. Buffer for losts in the traces.

 

So what if it was cherry picked?  You said it wasn't happening and it is.  Why would you even speculated that they changed the threshold?  At this point, you're just making stuff up. 

[ivan134]

3 hours ago, Lays said:

I mean ya can't really blame em, they're a company after all and the goal is to gain as much capital as possible while maintaining good customer satisfaction (In theory lol)

 

From what I've seen, node shrink doesn't directly correlate to perf increase, rather just less power draw.  Usually when people are shrinking nodes, they're also putting a brand new arch on it which seems to be where the extra perf comes from

The node shrink does bring performance increase though. A 314mm2 chip is giving the same performance as a 601mm2 chip. I guess i dont blame people for comparing the 1080 to the 980 ti though since Nvidia is now charging high end price for a card that isn't high end. 

 

As for the voltage, i think that might be a limit of a 14/16 nm node. The 980 ti should be compared to GP102 which should be a similarly sized chip, to see whether the voltage matters a much. Same goes for performance. 

[Vellinious]

Too many people bought the Fool's Edition.  They should have just waited for the custom boards.  The reference cards with the single 8 pin are going to be severely hampered by power limitations, unless they start soldering on add on boards to get around the physical limitations, they were never going to get very far with them anyway.

Have to admit....gives me a good laugh though.  

[Lays]

3 minutes ago, Vellinious said:

Too many people bought the Fool's Edition.  They should have just waited for the custom boards.  The reference cards with the single 8 pin are going to be severely hampered by power limitations, unless they start soldering on add on boards to get around the physical limitations, they were never going to get very far with them anyway.

Have to admit....gives me a good laugh though.  

I already said they tried soldering add on boards and the issue still persists, and a 1080 strix review also mentioned it being locked at 1.25v.

[done12many2]

3 hours ago, Vellinious said:

Too many people bought the Fool's Edition.  They should have just waited for the custom boards.  The reference cards with the single 8 pin are going to be severely hampered by power limitations, unless they start soldering on add on boards to get around the physical limitations, they were never going to get very far with them anyway.

Have to admit....gives me a good laugh though.  

 

Single 8-pin is not the limiting power factor with this card.  Actually, single 8-pin can deliver a great deal more power than folks think.

[Vellinious]

38 minutes ago, done12many2 said:

 

Single 8-pin is not the limiting power factor with this card.  Actually, single 8-pin can deliver a great deal more power than folks think.

lol, yeah....the power limit issues I've seen in a few of the overclock reviews must be a figment of my imagination.  

I've seen people set their bios to pull 200 watts from an 8 pin, have heard that you can go as high as 225, but....that'd rely on a really good PSU to deliver that much without tripping it up.  

We'll see where these fool's edition cards fall in the hierarchy when the custom pcb's start coming out with more available power.  I'd bet that the early adopters will either change up, or just choose to languish in mediocrity....just like nearly every other release since before I started messing with this stuff seriously back around Fermi.

But hey...if you've got the money to burn on a card that will likely get left behind within a couple months, then more power to ya.

[Vellinious]

47 minutes ago, Lays said:

I already said they tried soldering add on boards and the issue still persists, and a 1080 strix review also mentioned it being locked at 1.25v.

Gotta be some limitation that NVIDIA placed on the reference boards....NVIDIA loves their "safety" first mentality, ya know.  You can bet the aftermarket cards that are meant for LN2 cooling, won't have the same restrictions.

[done12many2]

3 hours ago, Vellinious said:

lol, yeah....the power limit issues I've seen in a few of the overclock reviews must be a figment of my imagination.  

I've seen people set their bios to pull 200 watts from an 8 pin, have heard that you can go as high as 225, but....that'd rely on a really good PSU to deliver that much without tripping it up.  

We'll see where these fool's edition cards fall in the hierarchy when the custom pcb's start coming out with more available power.  I'd bet that the early adopters will either change up, or just choose to languish in mediocrity....just like nearly every other release since before I started messing with this stuff seriously back around Fermi.

But hey...if you've got the money to burn on a card that will likely get left behind within a couple months, then more power to ya.

I'm not sure that you understood what my point was.  The issue you're talking about has nothing to do with the 8-pin connection you mentioned as the problem.

 

I do agree that things will get better, but I don't think dual 8-pin or 8+6 pin is going to improve things by itself.

 

As far as getting left behind, I don't put a whole lot of faith in that.  I did pretty damn good with some 980 Ti reference PCB cards on water with regards to benchmarking.

 

 

[Vellinious]

Well yeah, that limitation is what NVIDIA sets in the stock bios files.  What are they setting the stock power limit to after the 10% or 20% increase in power limit?  220?  230?  Having more power available to add in a bios is at the very heart of the problem.  It's funny...people said the same thing about the Maxwell GPUs.  "you'll never need that much power".  I needed 300 watts set in the bios as a minimum, or the 970s I played with would power limit throttle at the clocks I was running them at.  

As do others with reference boards....but the Matrix, OC Extreme, Classy, KPE, HOF edition cards are always just a little bit better.  And I'd be willing to bet, that with the single 8 pin on the FE cards this gen, that'll be even more pronounced.  We shall see....

[done12many2]

3 hours ago, Vellinious said:

 We shall see....

 

Agreed.

[Lays]

12 minutes ago, Vellinious said:

Gotta be some limitation that NVIDIA placed on the reference boards....NVIDIA loves their "safety" first mentality, ya know.  You can bet the aftermarket cards that are meant for LN2 cooling, won't have the same restrictions.

Did you not see the part where I mentioned it was a limit on the non reference Strix?  And with an E-Power soldered on, that bypasses all the VRM?  Or where I said even with a BIOS with all limitations removed, it still did it?

 

14 minutes ago, Vellinious said:

lol, yeah....the power limit issues I've seen in a few of the overclock reviews must be a figment of my imagination.  

I've seen people set their bios to pull 200 watts from an 8 pin, have heard that you can go as high as 225, but....that'd rely on a really good PSU to deliver that much without tripping it up.  

We'll see where these fool's edition cards fall in the hierarchy when the custom pcb's start coming out with more available power.  I'd bet that the early adopters will either change up, or just choose to languish in mediocrity....just like nearly every other release since before I started messing with this stuff seriously back around Fermi.

But hey...if you've got the money to burn on a card that will likely get left behind within a couple months, then more power to ya.

8 pin connectors aren't going to give out from 50w extra being pushed through them, when I benched my 980 ti on DICE at 1800 mhz 1.56v, I was pulling 700+ watts through 2 8 pin connectors, and it was perfectly fine for hours on end.  Legit all you need is a proper PSU, not some poop unit

[Vellinious]

1 minute ago, Lays said:

Did you not see the part where I mentioned it was a limit on the non reference Strix?  And with an E-Power soldered on, that bypasses all the VRM?  Or where I said even with a BIOS with all limitations removed, it still did it?

 

8 pin connectors aren't going to give out from 50w extra being pushed through them, when I benched my 980 ti on DICE at 1800 mhz 1.56v, I was pulling 700+ watts through 2 8 pin connectors, and it was perfectly fine for hours on end.  Legit all you need is a proper PSU, not some poop unit

Nope...didn't read that part, didn't care to.  Hate the STRIX cards, and skip over any post about them.

Yeah...that's what I said.  You'll need a damn good PSU to push that much out of an 8 pin without tripping it up.  lol

[done12many2]

3 hours ago, Lays said:

8 pin connectors aren't going to give out from 50w extra being pushed through them, when I benched my 980 ti on DICE at 1800 mhz 1.56v, I was pulling 700+ watts through 2 8 pin connectors, and it was perfectly fine for hours on end.  Legit all you need is a proper PSU, not some poop unit

 

Exactly.  This is my favorite part of the article you provided a link to in post #1.

 

 

STEP 3 – Power limit overrides

Often gamers and users are mistakenly referring to 6-pin or 8-pin MiniFit.JR connectors as 75W or 150W capable inputs. Nothing can be further from truth. These power levels are nothing but just way for NV determine how capable is used board hardware to deliver power. It’s purely imaginary number and have nothing to do with actual power taken from connector nor power input capability. Software and NV BIOS will handle GPU clocks and reduce voltages if measured power hitting programmed BIOS limit (which can be different value than 75/150W!).

 

So if we just change circuit to report lower power reading, this limitation will be lifted accordingly as well. Also to make sure we are not at any physical limit of power connector itself, check Molex 26-01-3116 specifications, which have specifications both 13A per contact (16AWG wire in small connector) to 8.5A/contact (18AWG wire). This means that using common 18AWG cable, 6-pin connector specified for 17A of current (3 contacts for +12V power, 2 contacts for GND return, one contact for detect). 8-pin have 25.5A current specification (3 contacts for +12V power, 3 contacts for GND return and 2 contacts for detection). This is 204W at +12.0V level or 306W for 8-pin accordingly. Now if somebody tells you that 6-pin can’t provide more than 75W, you know they don’t understand the topic well. It’s not the connector itself or cable limit the power, but active regulation of GPU/BIOS/Driver according to detection of used cables and preprogrammed limits. So how actual power measured?

[Lays]

2 minutes ago, Vellinious said:

Nope...didn't read that part, didn't care to.  Hate the STRIX cards, and skip over any post about them.

Yeah...that's what I said.  You'll need a damn good PSU to push that much out of an 8 pin without tripping it up.  lol

The PCB itself is fine, they just have a bit lackluster coolers on em it seems.  

[Curufinwe_wins]

1 minute ago, Lays said:

snip

In fairness to everyone involved... GM200 was a die literally larger than what was considered possible even a generation before it.

 

It was within 1% of the ABSOLUTE maximum die size allowable by TSMC's masks...

 

Even GP100 is much smaller than what 16nm will allow.