Gaelium (GAEL) mining

[nc50lc]

-snip-

You need to take into account the technical aspects of mining with older graphics cards. In my view, graphics cards should be replaced every 5–7 years, provided they haven't been used for hot algorithms and have been well-maintained.

This perfectly shows FP91G's experience on mining rigs.

For example, I had a few RX570 cards from when ETH mining was booming, had to underclock their VRAM frequency after a few years due to similar issues.
After they're decommissioned, I found out that most of the have developed "hidden issues" on their VRAM when used on something more heavy like gaming.
Works perfectly on idle or low resource usage but after reaching certain VRAM frequency, it freezes the machine.
Issues are similar to some of the cards so it must be its heavy VRAM usage during mining.

[FP91G]

-snip-

You need to take into account the technical aspects of mining with older graphics cards. In my view, graphics cards should be replaced every 5–7 years, provided they haven't been used for hot algorithms and have been well-maintained.

This perfectly shows FP91G's experience on mining rigs.

For example, I had a few RX570 cards from when ETH mining was booming, had to underclock their VRAM frequency after a few years due to similar issues.
After they're decommissioned, I found out that most of the have developed "hidden issues" on their VRAM when used on something more heavy like gaming.
Works perfectly on idle or low resource usage but after reaching certain VRAM frequency, it freezes the machine.
Issues are similar to some of the cards so it must be its heavy VRAM usage during mining.

I try not to wait for those moments to arrive; instead, I sell off my graphics cards. While it is possible to run them at lower settings, there’s already no profit to be made as it is, so using lower settings would just make mining completely unviable. If a graphics card can still pass gaming benchmarks, I sell it and purchase newer models.

[gaelium]

I try not to wait for those moments to arrive; instead, I sell off my graphics cards. While it is possible to run them at lower settings, there's already no profit to be made as it is, so using lower settings would just make mining completely unviable. If a graphics card can still pass gaming benchmarks, I sell it and purchase newer models.

Can confirm the VRAM thing, been through it.
I had 5 rigs with 6 RX470s each, same hardware everywhere, and still some were rock stable and others kept freezing.
Always figured one card or another had a problem... never found which one.

And yeah, when you've built and set up the rigs yourself it's better not to be too far away.
Basement is usually a good spot, the attic a lot less in my opinion.

Selling them while they still pass gaming benchmarks makes sense, that's probably how most of them end up with gamers anyway.
And it must be exactly the same with cards used for AI.
Since most of those will hit the second hand market one day, I think gamers are in for some nice crashes... 

On the other hand, running older cheap hardware also saves you from big investments, with GPU prices being what they are right now.

I'm messing around with a local AI myself and struggling to upgrade it, hardware prices are crazy, even second hand.

[FP91G]

I try not to wait for those moments to arrive; instead, I sell off my graphics cards. While it is possible to run them at lower settings, there's already no profit to be made as it is, so using lower settings would just make mining completely unviable. If a graphics card can still pass gaming benchmarks, I sell it and purchase newer models.

Can confirm the VRAM thing, been through it.
I had 5 rigs with 6 RX470s each, same hardware everywhere, and still some were rock stable and others kept freezing.
Always figured one card or another had a problem... never found which one.

And yeah, when you've built and set up the rigs yourself it's better not to be too far away.
Basement is usually a good spot, the attic a lot less in my opinion.

Selling them while they still pass gaming benchmarks makes sense, that's probably how most of them end up with gamers anyway.
And it must be exactly the same with cards used for AI.
Since most of those will hit the second hand market one day, I think gamers are in for some nice crashes... 

On the other hand, running older cheap hardware also saves you from big investments, with GPU prices being what they are right now.

I'm messing around with a local AI myself and struggling to upgrade it, hardware prices are crazy, even second hand.

If the logs don’t make it clear which graphics card is causing errors, you can identify it through indirect signs for instance, by checking the temperature -it will be lower or the fan speed -the fans will be spinning more slowly. Sometimes, however, the issue lies with the risers or the electrical contacts, I’ve occasionally had to spend an entire week right next to the mining rigs to figure it out.

[Akka 007]

Gaelium’s KawPoW algorithm is notorious for frying VRAM, so those "hidden issues" are likely permanent silicon degradation from excessive heat. Treat this project with extreme caution; it’s less about community building and more about burning through your hardware.

[nc50lc]

Gaelium’s KawPoW algorithm is notorious for frying VRAM, so those "hidden issues" are likely permanent silicon degradation from excessive heat. Treat this project with extreme caution; -snip-

If you're responding to my post with the same quote, that wasn't caused by Gaelium mining.
That was quite a long time ago during Ethereum mining era.

But yeah, it's applicable to this coin as well but that's pretty much unavoidable on such algo.
Better take FP91G's suggestion to sell them in time before those issues develop given that they already paid for themselves.

[FP91G]

Gaelium’s KawPoW algorithm is notorious for frying VRAM, so those "hidden issues" are likely permanent silicon degradation from excessive heat. Treat this project with extreme caution; it’s less about community building and more about burning through your hardware.

AMD Radeon RX 570 and RX 580 graphics cards as well as their 400-series counterparts, lack a hardware sensor for video memory temperature, you only see the GPU core temperature. Newer graphics cards have long included memory sensors, though software typically displays the highest temperature recorded among them.
Mining is a business, and if you choose profitable coins, a new graphics card will generally pay for itself before it breaks down.

[gaelium]

Funny timing for this discussion, I just lived it with my own hardware, and not on a mining rig.

I bought two used Tesla P40 24GB from the same seller in China for my local AI box, one has been perfect since day one.
The other threw an uncorrectable memory error in the middle of a job this week, the driver retired a page and the card needed a reset.
When I pulled its ECC counters it already had 173 corrected errors and 10 uncorrectable ones, its twin shows zero.

That card never mined anything on my side, it came that way, so I agree with nc50lc, VRAM wear is not about one coin.
Any memory heavy load, Ethash, KawPoW or AI work, will find the weak chips faster when memory runs hot for months.

FP91G's point about sensors is the real lesson, the P40 has ECC, so it tells you when its memory goes bad.
An RX470 just freezes and leaves you guessing which card it was, exactly like on my rigs.
With old consumer cards, keeping the memory cool and selling before trouble starts is still the best advice in this thread.

The seller asked for proof, so I sent the driver logs and the ECC report.
I'll post how the replacement goes...

[FP91G]

Funny timing for this discussion, I just lived it with my own hardware, and not on a mining rig.

I bought two used Tesla P40 24GB from the same seller in China for my local AI box, one has been perfect since day one.
The other threw an uncorrectable memory error in the middle of a job this week, the driver retired a page and the card needed a reset.
When I pulled its ECC counters it already had 173 corrected errors and 10 uncorrectable ones, its twin shows zero.

That card never mined anything on my side, it came that way, so I agree with nc50lc, VRAM wear is not about one coin.
Any memory heavy load, Ethash, KawPoW or AI work, will find the weak chips faster when memory runs hot for months.

FP91G's point about sensors is the real lesson, the P40 has ECC, so it tells you when its memory goes bad.
An RX470 just freezes and leaves you guessing which card it was, exactly like on my rigs.
With old consumer cards, keeping the memory cool and selling before trouble starts is still the best advice in this thread.

The seller asked for proof, so I sent the driver logs and the ECC report.
I'll post how the replacement goes...

The Tesla P40 is analogous to the GeForce GTX 1080 Ti, but with 24 GB of memory. What motherboard and processor are you using with these cards? Do you have many AI models?
Buying used graphics cards requires very thorough testing, and even then, there’s no guarantee that problems won’t arise a few months after the purchase.

[gaelium]

Yes, same Pascal chip as the 1080 Ti, but 24GB and no fan of its own.
It's made for server airflow, so I fitted each card with a small blower kit, quiet enough to live with at home.

They sit in an Intel S2600ST board with two Xeons and 128GB of ECC RAM, on Ubuntu with llama.cpp.
Mostly one big model, Qwen3 Coder 80B in Q4, split across both cards for a personal coding assistant I'm building, nothing commercial, just a hobby that got a bit out of hand.

You're right about testing, the funny part is that the bad card already had 9 uncorrectable errors and 9 retired pages when it arrived, it was all there in the counters.

It only crashed later, under a real load.
So my tip for anyone buying used datacenter cards, run nvidia-smi -q -d ECC the day you get them, before anything else.