I think something crazy happened here.

[alex843726351-web]

Hello,

I wanted to report an incident that occurred while I was using Odysseus for local AI inference.

I was running a Qwen model on my desktop system equipped with a Gigabyte RTX 4080 SUPER. During normal operation, I suddenly heard a popping sound from my computer, followed by a brief flash/light from the graphics card and a strong electrical odor. I immediately powered the system off and disconnected power.

After inspection, the smell appears to be localized to the graphics card itself. The GPU power connector does not appear to be melted or visibly damaged.

I am not claiming that Odysseus caused the hardware failure, as I do not have evidence of that. However, because the incident occurred while using the software, I wanted to report it in case there have been similar reports from other users or if there is any information that may be helpful.

Thank you for your time.

[AnnoDomine]

See: #3103

[alex843726351-web]

Get in contact to Gigabyte buddy. This the area dank means :) It is really not related to the project. 🫰

wait if the device light come from where I point that means it's not related to the project?

[AnnoDomine]

Exactly!
What I mean is: running an LLM just pushes your graphics card to 100% load. This heavy load most likely caused a weak spot in that physical area of your hardware to finally give up. The exact same thing could have happened while playing a very demanding game for hours.

I won't try to guess what exactly broke there, but it strongly points to a physical hardware failure of the board itself, not a software bug from this project.

But to be clear, please contact Gigabyte. They can give you a way better answer and hopefully help you out!

[alex843726351-web]

oh I see! by the way thank you all for helping! I appreciate it! also you AnnoDomine thank you for helping!

I let you guys know if something happen!

[dankxylese]

@dankxylese Ah, gotcha! So your comment was basically just a follow-up to the thermal panic in issue #3103?

Just out of curiosity and to get your take on this: :) I am running a 4070 Ti myself (bought in 2022). I’m currently training a custom-coded LLM, written 100% in Rust, and my GPU has been sitting at exactly 100% usage for 5 days straight now. No fires, no melted cables, no pops so far.

What do you say about that? :P

I'd say you're lucky with your card, that it has no defects and runs well, given the poor quality in the 4xxx series by Nvidia.

Other than that, you are describing the base any system should be able to pull. I'd even say, with good cooling, if your system is not overheating, you should be able to pull 100% load for multiple years non stop.

I used my 3070 for mining, and now LM work, along with my 5800x3D. And they've been running flawlessly for years, multiple days and weeks at 100% load. The only thing that gave out was the RAM, which started showing errors on memtest86. But I got it RMA'd and back to work.

[dankxylese]

That issue is nonsensical. There are multiple protection systems that kick in before a piece of hardware can fail. Multiple of those protections software wouldn't be able to change. I posted details under that issue.
The actual incident here is almost certainly unrelated to Odysseus. The RTX 4080 SUPER uses the 16-pin 12VHPWR connector, which had a documented and widely-reported failure mode in 2022–2024 - the adapter melts or arcs under sustained high load due to a marginal contact resistance issue (NVIDIA acknowledged this). A pop, flash, and electrical smell with no visible external connector damage is consistent with a capacitor failure on the PCB, which can happen during any sustained high-load workload — gaming, rendering, mining, or inference.
Running Qwen at sustained GPU load just happened to be what the hardware was doing when a pre-existing marginal component gave out. Same failure would have occurred running a game at max settings.

when you say that do you mean my Connector like the one should be like lighting or flaming? because it wasn't from there, I would upload a pic of my gpu and I will show you where I saw the flame came from

The connector is the most common flaw in the 4xxx and 5xxx series. Look up videos on GamersNexus, it was really well covered a year or two ago. But capacitors are also not good quality on select cards of those series, since they've switched them out mid production.

If you can't get your money back, give it to a repair shop to get checked out. If its a capacitor issue, you are looking at $5 replacement cost max, plus labor which could be $30-50 for the time to disassemble, diagnose, replace. Would be a relatively cheap fix given current market prices for getting a new GPU

[alteixeira20]

Thanks for your feedback and I'm sorry for what happen, hope it didn't cause long term damage, I'll be giving this problem priority!

[alex843726351-web]

yeah thank you! right now I got my old 1660 TI installed, and removed all the apps I was using, also I did make sure my PSU is fine, the only thing and I am not sure if it's gone 100% is the GPU and to be honest it might be gone sadly

[ryslan25500-cloud]

Congratulations, you are the first buyer of a new graphics card. Congratulations, you can pay for the purchase of a pewdiepie. + do not forget the moral penalty, + for the equipment, he has a lot of money.

[ryslan25500-cloud]

#3103 (comment) reed its help you. in yours situation

[ryslan25500-cloud]

#623
#607
#612

[ryslan25500-cloud]

1 diagnosis, send it to a gigabyte, 2 wait until people's video cards start to burn too, and the friction to sue in Japan. You will win the case with 100%.

[ryslan25500-cloud]

This is not a dead end, but this is the apogee, horror in the program code, do not go to negotiations, demand x2 1600USD + a collective lawyer, in Japan, let the Swedish pig cover everything.

[ryslan25500-cloud]

Consequently, you can consciously establish a yoke program code, and you can make money on it. See how he presents his billion-dollar product.

[ryslan25500-cloud]

[alex843726351-web]

hey so I did talk to Gigabyte and I got a respond from them, I do hope they really do something about it, if anyone knows does this usually means they will be helping me really, or we still not sure yet?

[alex843726351-web]

so is Odysseus really that dangerous, and it's really killing GPUs this way?

[ryslan25500-cloud]

2+2=

[alex843726351-web]

I'm sorry if I sounded clueless I wanted to do learn about how to self-host Ai I really didn't know that it might do something like that.

[ryslan25500-cloud]

Translator, you have a 100% winning case. File a lawsuit and it's done, I'm sure that in a pre-trial procedure. or through the court. In any case, I think you can pull out up to 5k USD calmly. with him. "AS IS" IS THE LATEST QUALITY STANDARD.

[ryslan25500-cloud]

He was chasing popularity, but he should come to prison, you could have a fire.

[alex843726351-web]

Hey, so I did talk to Gigabyte, and somehow they really agreed to replace my GPU. I will be getting a new 5080 Super since it's cheaper for them and better for me, so I think I am happy with the Change.

They did check it, and they said there is a capacitor failure, and the GPU was not abused in any way. They only asked me what I was doing, and I was worried, so I told them I was just using a normal heavy load Application.

They didn't ask any additional questions and just told me that I can go in a nearby Store it's like an hour drive from my house I give them my old GPU and invoice and they give me a new one.