AmpereGate! RTX 3080 instability

https://www.guru3d.com/news-story/g...ely-due-to-poscap-and-mlcc-configuration.html

During testing, we also re-ran the benchmarks, and it had offset effects that are close to zero, meaning at 100 FPS you'd perhaps see a 1 FPS differential, but that can be easily assigned to random anomalies as well. As to why there is so little performance decrease is simple, not many games trigger the GPU all the way the end of the spectrum at say 2050 MHz. That's isolated to very few titles as most games are GPU bound and hover in the 1900 MHz domain.
 
EVGA released another statement saying a nivida latest driver fixed the issue. Which begs to question if the driver now limits the boosting so the cards wont crash. Same bullshit with the 970 memory issue. They had to band aid it with a driver and pretend that fixed the issue.
 
Seeing reports that the issue isn't purely a capacitor issue but more completely: a stability issue with boost clock voltage ranges (capacitor design could smooth out the voltage swings more) that is probably addressed in the latest driver by adjusting the voltage on the boost ranges (not just "limiting" boost). Newer driver apparently also raises performance/FPS in many games. I.e. this may be a lot of furor and obsession with caps potentially over nothing.



This is just what I got out of it - not having checked the veracity of the details myself, but seeing other youtube videos saying "drivers" and finding some sense in what's being speculated here. Just more food for thought before this "amperegate" thing gets too out of control.
 
Seeing reports that the issue isn't purely a capacitor issue but more completely: a stability issue with boost clock voltage ranges (capacitor design could smooth out the voltage swings more) that is probably addressed in the latest driver by adjusting the voltage on the boost ranges (not just "limiting" boost). Newer driver apparently also raises performance/FPS in many games. I.e. this may be a lot of furor and obsession with caps potentially over nothing.
This sounds like it jives with the behavior der8auer noticed, where there's a big transient/spike in frequency right before crash - the GPU is stable around 1890/1905, he adds another 10Mhz and the boost overshoots to 2055Mhz for 1.5 seconds and crashes. And he noted that switching out the caps did *not* fix this behavior.

So from the sound of it they were able to smooth out that boosting behavior in-driver rather than requiring new vBios flashes. Going forward as new GPU stock is printed, perhaps a combination of the two.

As always, best to chill out and wait for things to develop. Same goes for new AMD cards when they come out and may have bleeding edge/launch issues.

1601407079452.png
 
Last edited:
EVGA released another statement saying a nivida latest driver fixed the issue. Which begs to question if the driver now limits the boosting so the cards wont crash. Same bullshit with the 970 memory issue. They had to band aid it with a driver and pretend that fixed the issue.
It is not the same thing and not a bandaid fix. Shot was rushed out and boasted higher then it should have been. You are still going to get higher then advertised boost clocks. Sure if the driver update cause the GPUs too boost under advertised clocks then it would of been a issue. It is not like you can't OC it yourself still.
 
It's not even factory OCs. These are stock clocked cards that are failing when they boost too high. In the video above the card is running TimeSpy fine at ~1950 and then fails when it instantaneously boosts to ~2055.
They are factory OC... boost speeds are 1710mhz... they are boosting to 2000mhz... how is that not a factory OC? If it's exceeding advertised specifications from the factory, then it's a factory OC. If it was limited to 1710mhz, then it wouldn't be a factory OC. Anything over that is above and beyond spec. This is why the solution of dropping boost down to 1900 would be legally fine as it was never advertised to be able to run more than 1710. Now, whether that would piss of their customers or if it's morally the right thing to do, well I'll leave that topic alone :p.


I had assumed that Nvidia giving AIBs drivers at the last second was intentional to avoid leaks. So not rushed but stupid.

As for launch inventory nobody knows how much stock was available. It would have sold out no matter what.
It was intentional... don't give them drivers that aren't working right == intentional :). Sorry, couldn't resist. This was the rumor I heard (they still had buggy drivers), but I don't have any way to confirm, so it could have been either/or or some of both.
 
EVGA released another statement saying a nivida latest driver fixed the issue. Which begs to question if the driver now limits the boosting so the cards wont crash. Same bullshit with the 970 memory issue. They had to band aid it with a driver and pretend that fixed the issue.
The only testing I saw (single source, so again, results may vary) was seeing 30mhz lower clocks (so very minimal) but 20w more power... so it seems the fix to the stability issue was to give it more juice (350w instead of 330w) and lower the speeds slightly. Will see if it's a permanent fix, but they saw no appreciable difference between back to back runs, so at least it seems it wasn't just pulling 300mhz off the top and calling it a day ;). Will wait for more reports though. That's putting it way up their in power and really skews any perf/watt tests comparing it to previous generations or upcoming AMD cards.
 
My biggest concern through all of this, is why did reviewers not say anything. IT seems some of them are now saying they couldn't get through runs of different games or w/e, but we didn't hear about the issues until consumers had them in hand. Where was all the reporting on this issues with review samples and why are they just admitting it now that they had issues rather than on the lift date? Nvidia is going to be for nvidia, AIBs are going to be for AIB's, but when did reviewers stop reporting on issues? Seems Kyle was about the only one man enough to give us the info straight up. (I didn't read all reviews, but just on glance, this wasn't an issue until it was brought up by multiple customers).
 
They are factory OC... boost speeds are 1710mhz... they are boosting to 2000mhz... how is that not a factory OC? If it's exceeding advertised specifications from the factory, then it's a factory OC. If it was limited to 1710mhz, then it wouldn't be a factory OC. Anything over that is above and beyond spec. This is why the solution of dropping boost down to 1900 would be legally fine as it was never advertised to be able to run more than 1710. Now, whether that would piss of their customers or if it's morally the right thing to do, well I'll leave that topic alone :p.

All Nvidia cards boost over the clock speed promised on the box. However, some cards are factory overlcocked so that their minimum promised clocks are higher. It doesn't make sense to conflate factory OC with boost as they are two different things.

It was intentional... don't give them drivers that aren't working right == intentional :). Sorry, couldn't resist. This was the rumor I heard (they still had buggy drivers), but I don't have any way to confirm, so it could have been either/or or some of both.

Not sure what you mean. Did you see the latest GN video? The drivers that Nvidia provided to AIB's only worked on 3dmark and furmark so AIBs had no opportunity to test their hardware on real games. This was an intentional move by Nvidia to limit leaks of game performance ahead of launch.
 
The real blame here is on nvidia for not wanting leaks to come out. They rushed everyone to get cards out.

This is why you can’t trust reviewers either. They had issues before the launch and didn’t say shit about the issues because nvidia wouldn’t send them cards in the future for throwing dirt on their name.

Like I said from the beginning this is a clock speed boost issue. Everyone got their panties in a bunch saying it’s the caps. If the boost clock is 1800mhz and the algorithm pushes it to 2000+ then they need to change it to not boost as high. It’s the same as a unstable overclock that many people here should of understood but they didn’t. They jumped on the reviewer band wagon and stirred up a ton of shit that didn’t need to happen.

People screaming the cards aren’t stable do have a point because this shouldn’t of happened. They should of tested the cards more with the boost. They could of locked the boost clock to max around 1850 MHz and this wouldn’t of happened and wouldn’t of stirred up all this shit. No one would of known about this til they went to manually overclock and it would crash then they would dial back and chalked this up to the cards running at max like the ryzen cpus people here love and all would be fine.

Ive enjoyed reading these threads. A lot of people here don’t know as much as they think they do and just jump on a so and so said this train. People wanting to return cards and get ones with different cap layouts have entertained me enough to get thru my F5 debacle of finding a card.
 
All Nvidia cards boost over the clock speed promised on the box. However, some cards are factory overlcocked so that their minimum promised clocks are higher. It doesn't make sense to conflate factory OC with boost as they are two different things.
Whether your increasing clocks over base or boost, it's still an OC. By your logic if I buy a 3080 and set my base clocks to 1200mhz and boost to 2500mhz, I underclocked my card because only the change to base clock matters?

Not sure what you mean. Did you see the latest GN video? The drivers that Nvidia provided to AIB's only worked on 3dmark and furmark so AIBs had no opportunity to test their hardware on real games. This was an intentional move by Nvidia to limit leaks of game performance ahead of launch.
No, sorry didn't see it, was going on previous rumors which may (apparently, ARE) out of date ;). I wonder how MSI and EVGA were able to get enough testing to find the issue if this is indeed true, unless they were seeing crashes in just these 2 benchmarks?
 
Whether your increasing clocks over base or boost, it's still an OC. By your logic if I buy a 3080 and set my base clocks to 1200mhz and boost to 2500mhz, I underclocked my card because only the change to base clock matters?

Nope I’m saying factory OC actually means something in the industry and it has nothing to do with boost.

No, sorry didn't see it, was going on previous rumors which may (apparently, ARE) out of date ;). I wonder how MSI and EVGA were able to get enough testing to find the issue if this is indeed true, unless they were seeing crashes in just these 2 benchmarks?

Probably but we know some AIBs shipped review cards with different cap configs than they shipped to retail. So we know they found the issue very late, likely after getting proper drivers.
 
That's what happens when you have long hair. It constantly gets in your face and you have to adjust it.
I had shoulder length hair for 5 years and was so glad when I finally cut it.
He should probably cut his, given his line of work. It's probably making him miserable and he may not even realize it.
 
+1 to the "why didnt any of the 100s of reviewers see this"

some people say games crash within 5 minutes consistently with the 1st driver.
 
Nope I’m saying factory OC actually means something in the industry and it has nothing to do with boost.
Guess we'll just have to disagree on this one. Anything boosting over factory clocks in my mind is an OC, otherwise it'd just be spec. Doesn't matter if it's a boost OC or a base OC. The 3080's are maintaining 1800+mhz on average, which means they are sustaining speeds > "boost clock"... this is a factory OC in my mind, whether you agree with the semantics, it's coming from the factory running faster than stock speeds. If you want to give it a new term, go for it, but I wouldn't call it non-OC either.

Probably but we know some AIBs shipped review cards with different cap configs than they shipped to retail. So we know they found the issue very late, likely after getting proper drivers.
Just realized I said MSI, I meant to say Asus and EVGA. But, yeah, they obviously noticed BEFORE the review date (as some reviewers said they were notified that EVGA cards were being fixed prior to benchmarking and asked to wait). So, we know the problem was known prior to the review embargo and launch, we don't know how long before, but it means NVIDIA was aware of the issue from more than 1 (decent brand) manufacturer having issues, yet they went ahead with launch on schedule. They knew (or at least should have known) that EVGA and ASUS woudn't be the last to find issues with their design, so they allowed this to happen one way or the other. I guess they figured the flak from crashing GPUs was better than launching late. They would rather let down their customers than their investors. The customers will forget about it in a month, so it's probably the better of the 2 choices really, even if it pisses of some people. As long as it's fixed before RDNA2 comes out, it really won't affect sales much given their limited stock anyways, it's still going to be sold out.
 
+1 to the "why didnt any of the 100s of reviewers see this"

some people say games crash within 5 minutes consistently with the 1st driver.

Same reason most of them didn't find driver issues with Navi. Small sample size. A few hundred cards (if that) vs several thousand.
 
Nver had issue like that with my Aorus 2080 Ti Waterforce on stock.I bought in april no had single crash
 
Guess we'll just have to disagree on this one. Anything boosting over factory clocks in my mind is an OC, otherwise it'd just be spec. Doesn't matter if it's a boost OC or a base OC. The 3080's are maintaining 1800+mhz on average, which means they are sustaining speeds > "boost clock"... this is a factory OC in my mind, whether you agree with the semantics, it's coming from the factory running faster than stock speeds. If you want to give it a new term, go for it, but I wouldn't call it non-OC either.


Sorry man, but you are wrong about this. The factory specs are the guaranteed boost clock the card will reach, but it will probably boost higher if you have good case cooling etc.

It's incorrect to say it's a factory Overclock just because it boosts higher than the specs.

There is a term for it, it's called Boost clock. And it's not a factory overclock. There is no semantics to argue, those are the facts.

You could argue that buying Factory Overclocked cards is a waste of money these days as even the base models boost pretty high.
 
Sorry man, but you are wrong about this. The factory specs are the guaranteed boost clock the card will reach, but it will probably boost higher if you have good case cooling etc.

It's incorrect to say it's a factory Overclock just because it boosts higher than the specs.

There is a term for it, it's called Boost clock. And it's not a factory overclock. There is no semantics to argue, those are the facts.

You could argue that buying Factory Overclocked cards is a waste of money these days as even the base models boost pretty high.
Yep, still disagree with you, but it's ok. So if a card was "OCd" to 1500mhz base clock, and boosted to the factory 1710mhz... And another card kept base clocks the same but boosted to 2000mhz... Your consider the first one overclocked and the second one factory clocked? Man these guys must be marketing wrong, they should just dump boost clocks back down to 1710 and just give a +50mhz base clock so they can sell all of their base models as overclocked models!! Ok, I get it you want to call it boost clocks except that already has a term and a number, which is officially 1710mhz. So we have a boost frequency (clock) that is > advertised boost frequency (clock)... Let's just call this greater than. So we have clocks greater than the advertised boost frequency (I think we are still in agreement to this point?). So we have this thing that we are fine calling a higher (or greater) frequency... But if we replace higher with over... And clock for frequency (which are synonymous), it's now deemed the wrong thing? In my mind, whether you agree or not, if something is running at a frequency higher than spec, then it's overclocked. Whether it is the base or boost clocks, its just a base OC vs a boost OC. If that doesn't make sense to you, that's ok, feel free to not use it and continue on in life, it really doesnt change anything for either one of us.
 
+1 to the "why didnt any of the 100s of reviewers see this"

some people say games crash within 5 minutes consistently with the 1st driver.

I have checked out paul's hardware 2nd video where he explains that he tested his cards at stock without crashes, after shit hit the fan he retested them with an overclock and then they all crashed albeit some with a higher clock then others. But ahything you overclock too far crashes so that's not unusual, they did however work fine at stock settings, but yeah it is also a small sample size.
 
Correct me if I'm wrong, but for those of us who have factory overclocked 3080s ordered, it seems counterproductive to pay extra for the O/C model if the "stable" Nvidia drivers back the clocks down.
 
That's not what's happening. We should wait for more comprehensive coverage but the driver is clamping down on sudden frequency spikes that exceed the current delivery of the card immediately after a low load situation. This is partly why power usage is actually up on average. OC cards should still perform better than stock cards if they provide the power delivery to back it up and not just OC with no improvements. Plus silicon lottery.
Just like x570 motherboards, I think component selection will turn out to help OCs on this card.
 
That's not what's happening. We should wait for more comprehensive coverage but the driver is clamping down on sudden frequency spikes that exceed the current delivery of the card immediately after a low load situation.
This.
 
I hope the issue has been resolved, would like to see tests dealing with any performance differences with the correction. Are the day 1 reviews valid for true performance?
 
Factory overclocks still apply and are covered by warranty, the only thing which got reduce here is peak Boost clocks for some users. Before/after benchmarks are generally within margin of error, so I don't get why people think the day 1 reviews are invalid.

Even without the capacitor issues, its not an exact science to see different boost clocks between driver versions.
 
Sure, but you pay for one thing and are getting something else. The reviews should definitely revisit to see if performance has dropped by a measurable amount.
 
Sure, but you pay for one thing and are getting something else. The reviews should definitely revisit to see if performance has dropped by a measurable amount.

What's the something else? You pay for a card that boosts up to 1710Mhz.
 
Correct me if I'm wrong, but for those of us who have factory overclocked 3080s ordered, it seems counterproductive to pay extra for the O/C model if the "stable" Nvidia drivers back the clocks down.

Hard to say, seems like some of the non OC cards have power limit locked down to 100% like the Gigabyte Eagle whereas the Asus TUF OC can go to 108%. Unfortunately I haven't seen two cards from one vendor compared to see if power limits differ between base and OC models.
 
Factory overclocks still apply and are covered by warranty, the only thing which got reduce here is peak Boost clocks for some users. Before/after benchmarks are generally within margin of error, so I don't get why people think the day 1 reviews are invalid.

Even without the capacitor issues, its not an exact science to see different boost clocks between driver versions.
No one said they are invalid, if the problem is fixed, did it change the performance? Are different cards affected more than others with performance changes if any? If you change the boost clock behavior that should also mean a retest to validate previous results are in order. Not saying it degraded the performance or improved performance but if it did or not should be determined is all.
 
What's the something else? You pay for a card that boosts up to 1710Mhz.
If you bought a card based off of benchmarks done and don't give a hoot about mhz speeds and now you have been informed the boost behavior has changed, first I would want to know if what I bought performed as shown previously. I paid for a card with an expected performance result presented on reviews and from the manufacturer. Looking at game benchmarks with previous boost speeds of 2000hmz+ which those results were posted all over the internet and now it is restricted to 1710mhz could be a very significant reduction in what I thought I paid for. What change was done to fix the problem and did it affect the performance significantly? It could have improved it or degrade it or no difference, who knows until it is tested. Now since I don't own an Ampere card, tried but not wasting anymore time with Nvidia in the near future, I just can't blindly accept day one results.

I know of no benchmarks, reviews etc. that were limited to 1710Mhz intentionally.
 
If you bought a card based off of benchmarks done and don't give a hoot about mhz speeds and now you have been informed the boost behavior has changed, first I would want to know if what I bought performed as shown previously. I paid for a card with an expected performance result presented on reviews and from the manufacturer. Looking at game benchmarks with previous boost speeds of 2000hmz+ which those results were posted all over the internet and now it is restricted to 1710mhz could be a very significant reduction in what I thought I paid for. What change was done to fix the problem and did it affect the performance significantly? It could have improved it or degrade it or no difference, who knows until it is tested. Now since I don't own an Ampere card, tried but not wasting anymore time with Nvidia in the near future, I just can't blindly accept day one results.

I know of no benchmarks, reviews etc. that were limited to 1710Mhz intentionally.

Ummm yeah but so far all the reports are that performance has not changed. Not sure what we're talking about.
 
From what I've been reading, the drivers lowered max boost speeds on some cards (not all apparently) and increased power consumption slightly. However, performance has stayed relatively the same or increased in some instances. People are hypothesizing that the cards that are not boosting as high are sustaining their higher boost clocks for longer which keeps overall performance the same or even better. And lastly, there is conversation that its not necessarily the high boost clocks that causes the cards to crash, but rather the very quick ramp up to max boost clocks that would cause it. As a result, certain capacitor arrangements couldn't keep up with this ramp up.

But who knows, we need actual professional, repeatable testing performed with variables under control.
 
Back
Top