Hacker News Reader

Comment by creato | Hacker News Reader

One of the issues people have mentioned in this thread is that cards sold by miners are often burnt out or nearly burnt out. In that case, it seems like cutting off miners from the secondary market would be an improvement for the secondary market.

View on HN

I someone who hasn't bought a /new/ card since a 980ti, I find this talk of 'crypto burned gpus' to be a bunch of hooey. 1080ti x3, 1660ti x2, 2080ti x2... (We keep three gamers in cards around here, those aren't 'oh noes, this one fried, get another)... not one of them has had an issue.

Nor did the six P104s I picked up for an research project. And those cards smelled.. funny. I'm pretty sure they'd never seen aircon, and had no identifiable oem markings that I could find, but afaik, they're still chugging along 18 months after /whatever/ was done to them in their former life.

Honestly, I wonder if 'gpu burn' wasn't nvidia's first attempt to mitigate the cashflow drain of the crypto-cast-off-market.. before they decided it was more convenient to blame the need for drastic measures on a supply chain issue of their own design. -sigh-

Its an issue most people don't worry about because they're running exactly 1 GPU, which they wanted because they wanted to game, and if it dies, it dies.

But Google, Amazon et al. are very familiar with the concept of hardware lifetimes. Internally, calculations for some algorithms are done based on MTBF values for components - i.e. the cost of doing a thing is a certain number of CPU-hours, and you know a CPU will give you X-hours before it's liable to be hardware-dead.

A second hand gaming GPU has a very different power-on profile to a mining GPU - a gamer is going to struggle to on average put more then 4 hours a day or so through a card, most likely less. A miner is going to have that thing powered on the whole time, permanently. Even if the GPU is fine, the DDR RAM has been aging and that stuff fails faster.

The short answer is: people are going to be asking for a discount based on the expected lifetime of ex-miner hardware, once we know what that is. If the answer is "well it might last 6 months or it might last till you replace your machine" then the price will now reflect that expectation and risk profile. Not to mention the dilution factor which will be actually failing hardware being sold off by miners without properly marking it - after all, you might as well let someone gamble that bit flips in the GPU only lead to graphics aberrations rather then bad hashes.

You and I know different gamers though. People who buy expensive gpus usually push it for 8 to 12 hours per day on average.

While the cards are used heavily, thermal cycling arguably is worse than running a card 24/7. These cards do not generally experience any thermal fluctuations and were probably kept within adequate temps

I don't know much about GPUs and am curious how this "burn out" happens, do you have a link for those of us who'd like to read more about this?

If your GPUs run too long hot, the board literally starts to melt (the plastics, solder, etc).

Miners care enough about power efficiency and longevity; they don't run cards super hot.

Most scale oriented mining rigs I've seen pack cards super close together, even if you run them at reduced frequencies thermal environment is not great. See https://us.v-cdn.net/5021640/uploads/editor/do/j86gsm9cxvl2.... for example.

Aside from a few top end models modern desktop GPUs are not designed for continuous operation at max load. When you combine than with higher temperature the lifespan of the cards is indeed reduced. The real question is how much this matters.

Latest Nvidia GPUs have large Tensor cores and RT cores but miners never use them. I wonder is it better for the situation.

on contrary, miners tend to run gpus undervolted and underclocked for efficiency. gamers tend to run gpus overclocked (and sometimes overvolted) for performance.

Too much heat causes flex and movement that can break solder traces, and makes it easier to let the magic smoke out.

Flex is caused when the temp changes. Mining cards stay at the same temp and don't move around a lot while overclocked gaming cards are getting flexed multiple times a day

There's three main ways GPUs wear out, silicon electromigration, electrolytic capacitor dryout, and fan bearing failure. And the first two happen much faster at higher temperatures. Temperatures are similar between gaming and mining, but miners run 24/7 compared to gaming 1-2 hrs/day, so the GPUs age like 10x faster. But still, GPU failures are pretty rare, I think the concern is overblown.

Miners undervolt and often underclock though: heat is their enemy, so they’re more particular about this than gamers are these days.

Totally true- I went through the process of manually overclocking my 2080ti for max hash rate and the best settings involved setting a power limit that is less than half the card’s 100% limit.

> Heat is their enemy

Sometimes literally: https://www.youtube.com/watch?v=fr8bp8a2QS4

"A Bitcoin Miner Heatstroked In His Sleep. This Is What Happened To His Organs."

They also sometimes do GPU BIOS modifications, which is a fun thing to discover once the card ends up on the second-hand market and you happen to buy one. Luckily the fix is to reflash the BIOS using a matching one from TechPowerUp GPU database, but this assumes that you even know that this is something to pay attention to.

Source: happened with an RX 560 that I bought second-hand. Driver installation failed in Windows 10 due to the modified GPU BIOS. Was fixed with a reflash of a stock GPU BIOS using atiflash.

Any idea what they modified?

I didn't check it, but if I had to guess, it might have to do something with allowing the card to run using settings that make it more suitable for mining. Someone more familiar with GPU mining can correct me here.

Usually memory timings, at least on my 5700 XT

I’ve purchased graphics cards from the used market, and used them for years. Never any problems. One of my go to cards is 5 years old … that’s when I bought it - I’m guessing it’s much older.

> electrolytic capacitor dryout

Has anyone spotted electrolytics on a card in the last ten years or so?

Uhhh every single one. Here's the Nvidia 3070 for example: https://www.techpowerup.com/review/nvidia-geforce-rtx-3070-f...

I don't recognize the manufacturer's mark on those, but they're almost certainly solid polymer capacitors and not electrolytic.

Anedoctal but my two last GPUs burnt out after a few years. Bought new GPUs from reputable suppliers. Never mined or overclocked besides factory overclock (which for most of the time I had disabled). I was playing a lot more than 1-2 hrs/day. I guess it was the worse of both worlds. Heavy usage + cycling.

Just normal wear, accelerated because it's always under load. The used cards are fine most of the time if the temperatures were stable under 80-90 degC.

The main cause of failure is the VRM, those components are the hottest, least reliable and least cooled.

Sadly, sometimes a blown transistor or capacitor can take out the whole GPU.

Reply on news.ycombinator.com