News for hackers. Stuff that matters.
Re:In 2019 RX470 were $70, RX580 $100 by npteljes on Tuesday September 20, 2022 @11:50AM (#32913411)
My logic would be that these cards are for consumer usage, and mining means 24/7 100% load. So, a year mining would be like 5 years of gaming, and even then you gamed 5 hours a day, every single day! So, mining might not destroy cards, but it's heavy use.

I think the used goods buyers' attitude is similar to those in the car market. Never mind the miles, if a car had heavy users, like a pizza company's delivery vehicle, it looks different in the market than a garage-stored family car used to visit the relatives.

Re:In 2019 RX470 were $70, RX580 $100 by hansvm on Tuesday September 20, 2022 @12:04PM (#32913646)
Counterpoint, heavy use mostly doesn't apply to solid-state devices (much how the expected lifespan of an incandescent bulb is independent of its previous usage), and to the extent it would matter the effect is probably dwarfed by not being stuck in a gaming PC full of dust and experiencing constant thermal shock as the games ramp up and down.

To the pizza car analogy, when I was in that game I put a hard 60k miles/yr on such vehicles, but I also regularly change the oil, keep great tires and brakes on them, replace bushings and other suspension components at the first sign of issue, etc... Partly because driving that many miles makes issues easy to spot due to familiarity and temporal coherence and partly because it is cheaper to fix many such things sooner rather than later after they've jostled something else out of place. Nobody ever asked if it was a pizza car. I listed it as well maintained and got top dollar each time because it was and their inspections showed that to be true.

Re:In 2019 RX470 were $70, RX580 $100 by falcolas on Tuesday September 20, 2022 @12:51PM (#32914455)
> heavy use mostly doesn't apply to solid-state devices

Sorry, but this is pretty much incorrect. Copper migrates, even in solid state devices, and those copper migrations cause breaks and shorts, which cause errors and failures. The smaller the scale, the faster this happens. There is absolutely a MTBF even for solid-state devices.

This entire slide deck is worth the perusal, if only for the "whoah, that actually happens?!" moments. But here's a relevant slide for this claim:

https://c3.ndc.nasa.gov/dashlink/static/media/other/Observed...

The money quote: "Commercial semiconductor road maps show component reliability timescales are being reduced to 5-7 years." (with attribution)

Another fascinating slide is this one, which shows how an IC can start to have a timing issue after being in service for a time.

https://c3.ndc.nasa.gov/dashlink/static/media/other/Observed...

Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @12:57PM (#32914564)
We've definitely seen 'memory go bad'. But, our failure rate is in the single digits, across a huge sample set across multiple years and in relatively extreme conditions (4 seasons, outdoors).

If you're buying something for like half of retail and rolling the dice on it, I think it is a fair bet. Even if you only get 1-2 more years out of a card that has been used before (for any purpose), then you've done well.

Re:In 2019 RX470 were $70, RX580 $100 by d_tr on Tuesday September 20, 2022 @01:47PM (#32915223)
Half of retail (which is still high) for a mining card with expired warranty? No thanks.
Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @02:07PM (#32915584)
Ok, $10, but you have to pay for shipping and min order is 10k. ;-)

For clarity, because @derac is confused below... this is a joke. Hence the winky face.

Re:In 2019 RX470 were $70, RX580 $100 by derac on Tuesday September 20, 2022 @02:01PM (#32915495)
Your incentive is to propagandize this point. you have 100k+ cards to sell on the market, by your own admission.
Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @02:06PM (#32915565)
It is amazing how crazy HN can get. I'm not propagandizing anything.

If we sell them for pennies on the dollar or e-waste them, doesn't matter to me.

We ROI'd these cards already.

Re:In 2019 RX470 were $70, RX580 $100 by seattle_spring on Tuesday September 20, 2022 @04:07PM (#32917091)
Dude it's nuts how often people sling the "shill" insult on HN. One of the laziest insults by far, and I'd guess incorrect very near 100% of the time.
Re:In 2019 RX470 were $70, RX580 $100 by derac on Tuesday September 20, 2022 @02:20PM (#32915781)
Whoa crazy to point out someone's conflict of interest. CrAzY

You are sitting on 10m+ of stock and you don't care about the money. OK bud guess I must be schizophrenic, damn.

You are literally trying to sell your cards to a sister comment...

"Ok, $10, but you have to pay for shipping and min order is 10k. ;-)"

Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @02:27PM (#32915884)
There is no conflict of interest, I'm not trying to sell cards here on HN.

You are literally making stuff up now. The 10m+ of stock is worthless if nobody would buy it. If anything, it'll cost us to e-waste it. So, the price ceiling, if we were to sell it, is simply the cost e-waste and moving it around.

That was a winky face... a joke. I was kidding because the OP was unhappy with the statement of half price, which I also made up as a random baseline.

Re:In 2019 RX470 were $70, RX580 $100 by derac on Tuesday September 20, 2022 @02:31PM (#32915927)
The goal is changing the public discourse. Thus the word propaganda. You wouldn't be propagandizing this point of view if you didn't think you would change people's minds. You are fully aware that this forum has an outsized influence on the tech/programming community.

Edit: to me it's clear that you are acting in bad faith since you called me crazy for pointing out an obvious conflict of interest.

Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @02:45PM (#32916101)
1. I didn't call you crazy, I said "it is amazing how crazy HN can get."

2. Tell me exactly how would it benefit me to change public discourse around the long term reliability of GPUs? The most I can think of is that it might drive up the prices of used cards a tiny bit in a very saturated market, but I doubt it. Especially since I'm just one data point and I'm only posting here on HN. Plus, as I said, I don't care what our cards sell for. Our baseline cost at this point, is to offset any price of e-waste and shipping.

3. I shared my findings about long term reliability because this is a tech community and I thought they'd be interested in it! OMG, I'm such a horrible person for thinking this way! That is sarcasm.

4. Allowing people to realize that mined on GPUs are actually ok (not all, but generally better than common thought) might mean that a few more find useful homes and not end up as e-waste. That's a good thing! I even suggested that people use it as a way to negotiate the price down!

Re:In 2019 RX470 were $70, RX580 $100 by derac on Tuesday September 20, 2022 @03:44PM (#32916804)
Fwiw, I don't think you're a bad person. I think you have a conflict of interest and they way you are responding to criticism is weird. You also handwaved the copper migration issue. Meh. Carry on. Personally, I took that as you calling me crazy. If that wasn't your intention, fair enough.
Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @04:48PM (#32917603)
Your criticism is to say that I'm spreading propaganda, which is untrue and it is fair for me to respond to that.

Now you're at it again with the criticism. I'm not handwaving anything. I'm stating what I saw after having the fairly rare experience of running 100k+ GPUs over multiple years in harsh environments.

Let's turn that around, do you have any experience with this at all? Or do you just criticize people?

It also doesn't benefit me in any way to share information like this. I'm only sharing this information now because we're pretty much done with this phase of the business, so I can talk about things a bit more openly, and it seems relevant to share on a thread about RX cards.

Re:In 2019 RX470 were $70, RX580 $100 by derac on Tuesday September 20, 2022 @05:15PM (#32917879)
I didn't say it was not ok to respond. Calling people crazy is the weird part.

You did handwave that, we can agree to disagree.

I didn't even say you are or aren't doing anything nefarious. I pointed out a conflict of interest (incentive to propagandize) and you responded by calling me crazy. Now you are saying "All I do is criticize". Boohoo to you sir.

Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Wednesday September 21, 2022 @10:44AM (#32925897)
Good luck with everything. I wish the best for you.
Re:In 2019 RX470 were $70, RX580 $100 by seattle_spring on Tuesday September 20, 2022 @04:20PM (#32917253)
"Assume good intent." This comment, as with all of your other ones in this chain, blatantly violate HN guidelines. Accusing people of "shilling" or "spreading propaganda" only makes this forum worse.
Re:In 2019 RX470 were $70, RX580 $100 by uhhyeahdude on Tuesday September 20, 2022 @01:08PM (#32914709)
Thanks for that slide deck! Many treasures to be found within.
Re:In 2019 RX470 were $70, RX580 $100 by falcolas on Tuesday September 20, 2022 @03:00PM (#32916291)
Any time I share it, I go through it again. It's so dang fun.
Re:In 2019 RX470 were $70, RX580 $100 by vineyardmike on Tuesday September 20, 2022 @12:44PM (#32914329)
Counter counter point, chips break down over time even though they’re solid state. The flow of electricity through a chip will slowly degrade the metal traces inside it. At the nanometer scale, it doesn’t take much to ruin a logic gate. This is called electromigration.

You can care for that car (or chip) all you want, but chips aren’t cars. It’s a slow process, and may not happen on the timescale of mining, but it happens.

https://en.wikipedia.org/wiki/Transistor_aging

https://en.wikipedia.org/wiki/Time-dependent_gate_oxide_brea...

Re:In 2019 RX470 were $70, RX580 $100 by AceJohnny2 on Tuesday September 20, 2022 @12:48PM (#32914402)
> heavy use mostly doesn't apply to solid-state devices

It kinda does. The effects are small, but increase as the fab node gets smaller.

https://en.wikipedia.org/wiki/Electromigration

Re:In 2019 RX470 were $70, RX580 $100 by ithinkso on Tuesday September 20, 2022 @01:26PM (#32914939)
> (much how the expected lifespan of an incandescent bulb is independent of its previous usage)

Either I'm reading this sentence incorrectly or there is something else I don't understand. How can that statement be true?

Re:In 2019 RX470 were $70, RX580 $100 by rowanG077 on Tuesday September 20, 2022 @03:20PM (#32916535)
It's possible if you have a completely flat failure rate. For every time unit spend it has x% failure rate. An example is a piece of rock on the side of the road. Everyday it has a small chance to be smashed by a car. But that chance doesn't increase or decrease over time. Unless external factors change

Whether it's true for light bulbs I don't know.

Re:In 2019 RX470 were $70, RX580 $100 by HenriTEL on Tuesday September 20, 2022 @04:57PM (#32917701)
I guess the bulb does not degrade over time, it just break at some point when some random events occur.
Re:In 2019 RX470 were $70, RX580 $100 by ithinkso on Tuesday September 20, 2022 @05:23PM (#32917983)
I don't think that's true, the filament evaporates (sublimes?, sublimates?) from the heat and eventually it is too thin to support itself and breaks. That's how I always understood the main failure mode of incadescent bulbs and that's why the above statement confuses me
Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @11:57AM (#32913524)
There is no such thing as 100% load in ETH GPU mining. The bottleneck is the memory controller, not the GPU itself. That's why 5+ year old cards are still more ROI efficient than the latest greatest cards. The only way to speed things up is to speed up the memory.

Other algos take advantage of more of the GPU side of things as well. ProgPoW, if it had passed, would have done that. From what I understand, RVN also does as well... and uses more power too.

Re:In 2019 RX470 were $70, RX580 $100 by garaetjjte on Tuesday September 20, 2022 @12:04PM (#32913654)
On the other hand 24/7 usage means no thermal cycling, probably eliminating failures caused by solder joint cracking.
Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @12:44PM (#32914340)
The machines reboot a lot for auto tuning. Plenty of thermal cycling that happens in all temperatures. Still didn't see any issues.

PSU's, on the other hand, are mostly hand soldered in China and have all sorts of issues.

Re:In 2019 RX470 were $70, RX580 $100 by Scramblejams on Tuesday September 20, 2022 @01:56PM (#32915364)
Who/what do you like for PSUs for operating in that harsh environment?
Re:In 2019 RX470 were $70, RX580 $100 by latchkey on Tuesday September 20, 2022 @02:50PM (#32916182)
We've run a ton of different PSU's. Everything from super cheap no-name chinese stuff to refurb/tested high end HP server class. None of them are super reliable.
Re:In 2019 RX470 were $70, RX580 $100 by PartiallyTyped on Tuesday September 20, 2022 @03:15PM (#32916478)
I trust mining GPUs a lot more than gamer owned GPUs tbh. Better ventilation, severe undervolting, not even maxing the gpu itself and the fans are not actually spinning that fast.
Re:In 2019 RX470 were $70, RX580 $100 by Gordonjcp on Tuesday September 20, 2022 @02:51PM (#32916193)
The garage-stored car will be a bag of shite, though. You'd be far better going for the delivery vehicle.
Re:In 2019 RX470 were $70, RX580 $100 by yieldcrv on Tuesday September 20, 2022 @01:13PM (#32914766)
undervolted use because not every processing component needs to be used

very different than a car

but valid questions to ask to whoever you are buying from used