NVIDIA recognizes the RTX 2080Ti issues

[bitsandpieces]

https://forums.geforce.com/default/topic/1079930/geforce-rtx-20-series/rtx-2080-ti-founders-edition-contact-us/

 

Quote

Limited test escapes from early boards caused the issues some customers have experienced with RTX 2080 Ti Founders Edition.

We stand ready to help any customers who are experiencing problems.

 

The "issues" in question are black screens, OS and system crashes/hangs, artifacting

They do not, however, point out the exact nature of the defect with these cards, since many have speculated it could be related to the new GDDR6 to it being way to close the the VRM

To note, NVIDIA only takes responsibility for their Founders Edition cards; the responsibility for AIB partner cards are in their respective corners, and I suspect the AIB partners are not keen with dealing with this situation especially after the debacles in the tech press in recent times

 

Further clarification:

"Test Escapes" refers to components on the PCB like resistors or capacitors that are tested offsite and can be of marginal quality and not known until after the fact

Other cause for is is malfunctioning test equipment

https://www.evaluationengineering.com/avoiding-test-escapes-when-is-good-really-good

Edited November 15, 2018 by bitsandpieces

[RejZoR]

So, only reference designed cards have this issue and not custom PCB designed ones?

[bitsandpieces]

1 minute ago, RejZoR said:

So, only reference designed cards have this issue and not custom PCB designed ones?

At the moment it seems so, but there are AIB cards who followed the reference PCB design in their respective releases

[RejZoR]

That's what I meant. Founders edition cards and those following reference design.

[bitsandpieces]

14 minutes ago, RejZoR said:

That's what I meant. Founders edition cards and those following reference design.

If it's the design, then it's a problem for everyone

If it's the components used, it can vary since AIBs have their own chain of supply

[CTR640]

Founders Editions Founded black screens, OS and system crashes/hangs, artifacting and more troubles to be Found.

[i_build_nanosuits]

personally i think the problem is GDDR6 memory itself because on youtube there are people uploading videos of cards having issues and it seem to be all over the place...like all kind of cards from entry level to top end seem to get it...i might be wrong but i wouldn't touch an RTX card with a 10 foot pole at the moment personally.

[cj09beira]

1 minute ago, i_build_nanosuits said:

personally i think the problem is GDDR6 memory itself because on youtube there are people uploading videos of cards having issues and it seem to be all over the place...like all kind of cards from entry level to top end seem to get it...i might be wrong but i wouldn't touch an RTX card with a 10 foot pole at the moment personally.

it could be many things, it could be instable voltages it could be wrong memory timings, etc etc

[i_build_nanosuits]

2 minutes ago, cj09beira said:

it could be many things, it could be instable voltages it could be wrong memory timings, etc etc

yes indeed exactly that's what i mean't by ''GDDR6 memory itself''...something wrong with it...can be many things.

[Dabombinable]

Notice how it was primarily reference PCB 2080ti having problems?

2 hours ago, i_build_nanosuits said:

personally i think the problem is GDDR6 memory itself because on youtube there are people uploading videos of cards having issues and it seem to be all over the place...like all kind of cards from entry level to top end seem to get it...i might be wrong but i wouldn't touch an RTX card with a 10 foot pole at the moment personally.

If the cards really needed improved memory bandwidth, Nvidia could have just used an even larger bus. Either way though people aren't that smart if they buy any of the RTX series, as all we are currently seeing is the cut down GPU. Instead of the 12GB, 384bit bus Titan version.

[mr moose]

What about all the reports of people fixing the issue by under clocking the memory? Is this the same issue or a new one?

 

 

[D13H4RD]

Just now, mr moose said:

What about all the reports of people fixing the issue by under clocking the memory? Is this the same issue or a new one?

 

 

That probably adds more info to the GDDR6 theory 

[TechyBen]

Just now, mr moose said:

What about all the reports of people fixing the issue by under clocking the memory? Is this the same issue or a new one?

 

 

IIRC Gamers Nexus found 2 issues. One was driver related, one was memory. To the latter, the memory is probably a hardware fault, and a firmware/driver might underclock it.  It just depends on how they offer the "fix", hardware or software.

[Amazonsucks]

4 hours ago, Dabombinable said:

Notice how it was primarily reference PCB 2080ti having problems?

If the cards really needed improved memory bandwidth, Nvidia could have just used an even larger bus. Either way though people aren't that smart if they buy any of the RTX series, as all we are currently seeing is the cut down GPU. Instead of the 12GB, 384bit bus Titan version.

Its not really a cut down GPU unless you consider all the xx4 GPUs cut down.

 

The xx4 GPUs have had a 256bit memory bus width for ages now. Nvidia changed the fact that the x80 card was based on the bigger xx0 GPU and began alternating their releases. It used to be xx0 chip is released as x80 card, x70 card and the xx4 chip is released as the x60 or x60ti.

 

The 680 used the GK104, which was really the successor to the chip used in the 560ti and 460. They got away with it because the GK104 was a new process node and new architecture at the same time.

 

780 then used the GK110 which was their high end GPU(ECC registers, 384bit bus, lots of double precision).

 

980 was the GM204, so named because the 750 used the GM107 die and therefore the 980 was part of the second generation Maxwell. It was, like the 680, the small chip as the x80 card though.

 

980ti was a totally new chip(GM200) which was a "big" GPU, but lacked any double precision capability or ECC registers.

 

Shortly after, the GP100 was released, but nowadays the true big chips GP100 and GV100 use HBM and have 1:2 DP with ECC registers. The bigger consumer chip became the xx2 chip(GP102) but its actually the successor to the GM200 and represents a bigger separation between their desktop and datacenter parts. Confusingly, the GM200 was used in the 980ti instead of making it part of the 1000 series.

 

With the 1000 series GPUs they used the GP104 as 1080 and GP102 as the 1080ti.

 

The only way to get the real big chip now is the Titan V or a Quadro V100 or GP100. GP100 was the first big chip not used in any way in a consumer part too.

 

So really the Titan Xp is also a cut down chip with no DP or ECC registers, even though it has the 384bit bus and it remains to be seen what the real big chip will look like. Right now its just called Volta Next and i assume it will be 7nm, use HBM3 and be more similar to GV100 than anything Turing released so far.

[BuckGup]

Hahaha what a shit show

[Tech Enthusiast]

4 hours ago, TechyBen said:

IIRC Gamers Nexus found 2 issues. One was driver related, one was memory. To the latter, the memory is probably a hardware fault, and a firmware/driver might underclock it.  It just depends on how they offer the "fix", hardware or software.

Can you link the memory error they found? Because i don't recall Steve talking about that at all.

The issue linked in this topic is basically Fox screwing over the QA step. They cheaped out on it and a few broken GPUs slipped trough.

 

That is also the reason why only FE are affected and nothing else, unless a board partner actually bought a preassembled FE and relabeled it. Usually board partner don't do that, so even reference designs from them are fine. Again, unless they did indeed buy a fully assembled FE and rebranded it. Which, judging from the RMA rates, noone did. 0.1 - 1.4% RMA rate is exeptionally low for a GPU, but those figures don't include the FE.

 

So, the issue is all with the FE only. The random broken cards from Partners are down to common issues that happen for every product in the universe. Just at a tiny 0.1-1.4% rate.

[Misanthrope]

I'd like to get confirmation on what exactly the issue is to begin with because there's very different things that can be happening: cost saving measures on a product with such a high premium and high performance than none should be spared, or maybe an anticipated release though that one seems unlikely as there's seemingly no way whatsoever for AMD to even respond to the high end segment at this point.

 

I think however that if this was something that could have been foreseen with more time (to either get the tech right, test it more or so far as to delay it to a smaller node) then it obviously means that Nvidia might be getting too aggressive with the investors talk and wanted to commit to just getting Ray Tracing out of the door when it might be the case that the current tech they have for it was pushed a bridge too far for it to handle it without compromising Quality Assurance.

[TechyBen]

34 minutes ago, Tech Enthusiast said:

Can you link the memory error they found? Because i don't recall Steve talking about that at all.

The issue linked in this topic is basically Fox screwing over the QA step. They cheaped out on it and a few broken GPUs slipped trough.

 

That is also the reason why only FE are affected and nothing else, unless a board partner actually bought a preassembled FE and relabeled it. Usually board partner don't do that, so even reference designs from them are fine. Again, unless they did indeed buy a fully assembled FE and rebranded it. Which, judging from the RMA rates, noone did. 0.1 - 1.4% RMA rate is exeptionally low for a GPU, but those figures don't include the FE.

 

So, the issue is all with the FE only. The random broken cards from Partners are down to common issues that happen for every product in the universe. Just at a tiny 0.1-1.4% rate.

The screen error they saw in their live stream testing (like this week vids list), suggests memory is at fault. No idea if the tester fried them, or just missed flagging them for "faulty" or to "refurb".

[rawrdaysgoby]

Nvda doesn't even do quality assurance anymore for their consumer cards anymore This is basically the same with intel and amd cpu situation...  They got lazy! 

[Tech Enthusiast]

5 minutes ago, rawrdaysgoby said:

Nvda doesn't even do quality assurance anymore for their consumer cards anymore This is basically the same with intel and amd cpu situation...  They got lazy! 

Foxconn does it, after assembly. 

That is exactly the step that went wrong on the first batch. At least to... you know... sources. If that is indeed the case, we will never know. But that is the story now and we can either buy it, or not. 

After all is said and done, all hints point toward it being a true story, but who knows.

With the very low RMA rates, it has to be a very isolated issue for FEs, otherwise the RMA rates would be 10 times higher then they are.

[Dabombinable]

5 hours ago, Amazonsucks said:

Its not really a cut down GPU unless you consider all the xx4 GPUs cut down.

 

The xx4 GPUs have had a 256bit memory bus width for ages now. Nvidia changed the fact that the x80 card was based on the bigger xx0 GPU and began alternating their releases. It used to be xx0 chip is released as x80 card, x70 card and the xx4 chip is released as the x60 or x60ti.

 

The 680 used the GK104, which was really the successor to the chip used in the 560ti and 460. They got away with it because the GK104 was a new process node and new architecture at the same time.

 

780 then used the GK110 which was their high end GPU(ECC registers, 384bit bus, lots of double precision).

 

980 was the GM204, so named because the 750 used the GM107 die and therefore the 980 was part of the second generation Maxwell. It was, like the 680, the small chip as the x80 card though.

 

980ti was a totally new chip(GM200) which was a "big" GPU, but lacked any double precision capability or ECC registers.

 

Shortly after, the GP100 was released, but nowadays the true big chips GP100 and GV100 use HBM and have 1:2 DP with ECC registers. The bigger consumer chip became the xx2 chip(GP102) but its actually the successor to the GM200 and represents a bigger separation between their desktop and datacenter parts. Confusingly, the GM200 was used in the 980ti instead of making it part of the 1000 series.

 

With the 1000 series GPUs they used the GP104 as 1080 and GP102 as the 1080ti.

 

The only way to get the real big chip now is the Titan V or a Quadro V100 or GP100. GP100 was the first big chip not used in any way in a consumer part too.

 

So really the Titan Xp is also a cut down chip with no DP or ECC registers, even though it has the 384bit bus and it remains to be seen what the real big chip will look like. Right now its just called Volta Next and i assume it will be 7nm, use HBM3 and be more similar to GV100 than anything Turing released so far.

GTX 1080ti (cut down GP102): 354bit bus, 11GB GDDR5X

Titan Xp (full GP102): 386bit bus, 12GB GDDR5X

RTX 2080ti (cut down TU102): 354bit bus, 11GB GDDR6

Quadro RTX 8000 (full TU102): 386bit BUS, 12GB GDDR6

 

The 2080ti is cut down. And the Titan Xp uses the full GP102. Nvidia does not cut down a GPU then change the moniker. GP100 is designed for HBM2, GP102 for GDDR5/5X.

[Zandvliet]

There aren't reports detailing the same problems with the 2080 correct? I have a 2080 and haven't had any problems so far.