Sciences proves mining kills cards

[Strategic]

1 minute ago, leadeater said:

Well I'm afraid that their testing and sample data is likely filled with bad data as the literal hundreds of thousands, no probably millions, of server GPUs proves this research data wrong in real life.

It is peer reviewed, and degradation is not equal to gpu dying. 

[SansVarnic]

7 minutes ago, Strategic said:

"As to what degree is under argument." Not really as shown in youtube video referencing the research, at 350  (1157,407 days) kelvin at already degrades significantly faster than than at 300 kelvin as 400 kelvin compared to 350 kelving ( 100000 seconds = 116 days)

 

I'm going to say this study is about as accurate as my 4 year old doing math.

I bought my GTX 1080 and ran F@H and BOINC on it for 24/7 16months out. The card ran much harder than it did when I mined on it. The card still works just as well as the day I bought it.

The metric they are using is just not viable.

Just saying.

 

Also you still need to clean up your original post.

[leadeater]

3 minutes ago, Strategic said:

It is peer reviewed, and degradation is not equal to gpu dying. 

Peer reviewed publications can and have still be incorrect and have had to issue retractions. This is sadly not as uncommon as it should be, still very uncommon but it still happens.

 

Edit:

Now science being wrong isn't a problem in itself, that whole point of science is to explore. But we and everyone other scientist will have to apply critical thinking to anything that is published and hopefully if the subject is important enough verified by other researchers conducting their own tests to further validate the finding.

 

Things really only become accepted fact after multiple independent studies shows the same thing

[Moonzy]

6 minutes ago, Frostbobomb said:

Hmmm, anyway for me to check the Tjunction? Or is that not possible on EVGA?

Is there anything in HWinfo64 about memory junction temp? If not then I'm not too sure

EVGA might not be reporting or pulling from the sensors if it's not present, which is kind of weird and concerning, because that's the most accurate way to measure temps

 

8 minutes ago, leadeater said:

the common misconception is that cards will die in 3 months like this source claims which is SOOOOO clearly false.

I can debunk this

 

I bought my GPUs around December and January

It is now... Mid April, none of them died (yet, -touch wood-)

 

Though one of the GPU fans are whining ever so slightly, I'll probably RMA it after stock stabilise and they won't offer a refund instead of repair/replacement

[Frostbobomb]

LOL if cards really died 3 months, who would ever make ROI?

[Moonzy]

6 minutes ago, Frostbobomb said:

LOL if cards really died 3 months, who would ever make ROI?

GPU I bought in December took ~3 months to ROI

 

because

1) cards prices weren't insane (yet)

2) the profitability from December to february was nuts

[Strategic]

21 minutes ago, leadeater said:

the common misconception is that cards will die in 3 months like this source claims which is SOOOOO clearly false.

Rewatched the video part, but he never stated such thing. He said to start throwing memory errors (elevated retention time degradation) within 4 month of usage, is exactly what he was talking about the memory retention. Meaning that it will get slower due to the errors, which certainly can affect the lifespan of the card. 

[leadeater]

22 minutes ago, Strategic said:

Rewatched the video part, but he never stated such thing. He said to start throwing memory errors (elevated retention time degradation) within 4 month of usage, is exactly what he was talking about the memory retention. Meaning that it will get slower due to the errors, which certainly can affect the lifespan of the card. 

The problem there is a lot of things, first I'll start off with what is being tested. Now you'll have to excusing any errors in this evaluation as he's jumping all over the place between research papers and some I cannot be bothered logging in to work to access as the full papers are not public.

 

In one of them he is referencing research that has being done on DDR1, DDR2 and FBDIMM, obviously all very old technology. These older memory technologies use higher operating voltages to modern memory and do actually just not last as long from my experience. Newer memory is also manufactured using better processes, how much this affects typical life span I don't know as even DDR1 and DDR2 lifespans are very long, I still have memory of this age working.

 

When it comes to GDDR memory this is actually a different memory type with different operational and functional parameters in mind, one of those being generally higher operating temperatures compared to regular DRAM.

 

Another problem is some of the data being presented is worst case, which is very problematic if you use this incorrectly. The other problem with worst case is without a huge sample size and knowing exactly how the worst case sample points were chosen we don't know what the actual underlying cause of the degradation is.

 

Then to bring it to the practical real world application mining specifically is highly sensitive to memory errors and if they happen the current job fails and is discarded, these errors are logged in the mining software. If errors really were a common thing and to the extent these research papers are being interpreted as then like I mentioned mining would not be a thing as the hardware would become non functional for the purpose before ROI.

 

Basically good science can be used badly.

 

Edit:

Not that I fundamentally disagree with the warning specifically being about RTX 30 series cards with GDDR6X as the memory does run much hotter, the problem there is this appears to be normal not abnormal (to some extent) so we have no idea how GDDR6X acts over longer term. It's just too new and I highly doubt research in much older memory technology is all that directly applicable.

[Strategic]

13 minutes ago, leadeater said:

The problem there is a lot of things, first it'll start off with what is being tested. Now you'll have to excusing any errors in this evaluation as he's jumping all over the place between research papers and some I cannot be bothered logging in to work to access as the full papers are not public.

 

In one of them he is referencing research that has being done on DDR1, DDR2 and FBDIMM, obviously all very old technology. These older memory technologies use higher operating voltages to modern memory and do actually just not last as long from my experience. Newer memory is also manufactured using better processes, how much this affects typical life span I don't know as even DDR1 and DDR2 lifespans are very long, I still have memory of this age working.

 

When it comes to GDDR memory this is actually a different memory type with different operational and functional parameters in mind, one of those being generally higher operating temperatures compared to regular DRAM.

 

Another problem is some of the data being presented is worst case, which is very problematic if you use this incorrectly. The other problem with worst case is without a huge sample size and knowing exactly how the worst case sample points were chosen we don't know that actual underlying cause of the degradation.

 

Then to bring it to the practical real world application mining specifically is highly sensitive to memory errors and if they happen the current job fails and is discarded, these errors are logged in the mining software. If errors really were a common thing and to the extent these research papers are being interpreted as then like a mentioned mining would not be a thing as the hardware would become non functional for the purpose before ROI.

 

Actually higher DRAM density is more sensitive to heat, that's why it even plays more. Same with 7nm CPU's degrading faster due to higher voltage and heat (my 3600 degraded 0.1 ghz within 1 year from 4.3ghz allcore to 4.2 ghz at higher voltage). That's the reason the research is even more relevant because they've found high density DRAM cells to be degrading faster.

 

Quote from abstract: "Abstract: The band-to-defect tunneling (BDT) induced junction leakage current of high density DRAM cell transistors under off-state bias-temperature (B-T) stress was studied in detail for the first time. It was found that the BDT leakage current is most critical for limiting the cell transistor scaling. The new off-state B-T stress was proven to be a very effective reliability assessment tool for leakage current degradation of the DRAM cell transistor. It was also found to be useful for assessing reliability degradation of future high density DRAMs." -https://ieeexplore.ieee.org/document/922872

 

But thing is, CPU's generally survive way longer than anybody will probably use the CPU. So if AMD goes 5nm the cpu will prolly drop from 14 years to 13 years life expectancy same with GPU chip, so that's  no problem. But VRAM on videocards seem to die way quicker, whereas the GPU chip is pretty much the last thing that will die on a gpu. It simply will then not turn on, whereas with VRAM problems you get these weird artefacts on your screen as VRAMs are like chiplets so if one dies gpu can still work albeit with artefacts.

 

--------------

 

But I appreciate your take, otherwise I wouldn't have had the chance explaining my view on it.

 

[leadeater]

10 minutes ago, Strategic said:

Same with 7nm CPU's degrading faster due to higher voltage and heat (my 3600 degraded 0.1 ghz within 1 year from 4.3ghz allcore to 4.2 ghz at higher voltage). That's the reason the research is even more relevant because they've found high density DRAM cells to be degrading faster.

But you have no idea what the actual cause is, could very well be external to the CPU and be related to VRM and input voltage signal quality. 

 

10 minutes ago, Strategic said:

Actually higher DRAM density is more sensitive to heat, that's why it even plays more.

Such is why operating voltage is lower and the junction temperatures delta to case temperature is lower than older memory technology that uses 1.5V or 1.8V rather than 1.25V or 1.35V. The entire way memory operates has changed significantly to account for all these factors. This is also why DDR5 is incorporating ECC function in to the specifications itself so memory density can be increased. Each memory generation incorporates changes to allow it to work and last as long or longer than what came before.

 

10 minutes ago, Strategic said:

But VRAM on videocards seem to die way quicker

But do they? We have many GPUs in 1U servers running very long time and have not seen abnormal failures at all. And that is the case for many thousands of other server operators. Cards are not dying before they become functionally useless, not at an abnormal rate.

 

When it comes to GPUs prematurely dying this is far more likely to be related to AIB OC profiles operating the memory above rated voltages to achieve increased performance. Voltage kills much faster than heat, so muuuch faster.

[leadeater]

13 minutes ago, Strategic said:

It was also found to be useful for assessing reliability degradation of future high density DRAMs." -https://ieeexplore.ieee.org/document/922872

Just FYI you know that is a 2001 paper?

[Strategic]

 

44 minutes ago, leadeater said:

Such is why operating voltage is lower and the junction temperatures delta to case temperature is lower than older memory technology that uses 1.5V or 1.8V rather than 1.25V or 1.35V. The entire way memory operates has changed significantly to account for all these factors. This is also why DDR5 is incorporating ECC function in to the specifications itself so memory density can be increased. Each memory generation incorporates changes to allow it to work and last as long or longer than what came before.

Well as for the 3000 series, we've yet to find out as it appears for example 3090 to run at 100 degrees when mining and 86 degrees without overclocking memory. Older GPU seem not to have that hot VRAM temps although they still die. I've had 3 gpu's die on me past 14 years (geforce 8800 2x and gtx 280 , they all had lines (sign of vram not working properly) on my screen and couldn't fix them by putting my last one in an oven (yes that was one of the solutions you could try to fix vram). My current MSI 1080 Ti (3+ years old) is already doing with occasional lines  if I overclock my memory with +50 mhz (and coil whines now) probably because I have been mining past 1,5 month and earned €200 with it (it could handle +200 mhz before). So I stopped mining and run my memory at stock now, hoping it will survive otherwise long enough for gpu prices to recover xd. 

 

 

-----------------

 

My initial point of the topic was that Linus has made several video's stating that miners are not a worse if not better buy for second hand, because they run their gpu's at lower temperature. But he ignores the fact that VRAM temperatures are not lowered, they're actually overclocked and run higher than stock. For example 3090 without overclocking memory it should run around 86 degrees or lower. Whereas miners that want increase their mining efficiency overclock their memory as that is the most important thing to overclock for mining.

 

 

 

 

[leadeater]

18 minutes ago, Strategic said:

couldn't fix them by putting my last one in an oven (yes that was one of the solutions you could try to fix vram).

That's disappointing, my 8800 GTS was fixed this way, still works today (not that I use it lol).

[IkeaGnome]

11 hours ago, Strategic said:

last one in an oven (yes that was one of the solutions you could try to fix vram).

You're not actually doing that to "fix" the vram. Heat cycles will cause solder joints to get stiff and eventually start to crack and break. Baking the GPU melts the solder to fix cracks and breaks in the solder. 

Think about it. If heat kills VRAM, how does higher heat fix VRAM?

[Zulu-Cheese-Alpha-IV]

Sorry, what science? What peer reviewed council gave way to this objectively bad argument and premise?

[Mark Kaine]

Well, I don't think this is "news" as in actual new information,  and in my opinion the biggest problem is people not knowing what they're doing, overclocking to get more $/t ,not having adequate cooling and for good measure put a lot of voltage and use dodgy "mining bios".

 

Of course it will degrade or outright kill gpus (which is why I'd never buy used gpus, from strangers,  because especially ebay & co is where those usually end up)

 

33 minutes ago, Zulu-Cheese-Alpha-IV said:

Sorry, what science?

"Youtube"

 

 

[Mick Naughty]

If the card lasts years mining, it isnt taking anything off its life span. As it lived through it already.