mia
mia@movsw.0x0.st
chromium moment
i have 30 GB video RAM on this laptop via GTT and it wanted all of it
3
1
2
1
0
2
johnny peligro
mischievoustomato@0.5dollah.click@mia how does that even happen? I've never run into this but I use Intel
1
0
0
johnny peligro
mischievoustomato@0.5dollah.click
0
0
0
@mischievoustomato some websites can do this via, i’m guessing, CSS animations, video playback, and inefficient shadow DOM updates. but i have not figured out the exact mechanism. the same thing on my intel laptop would cause the GPU to lag/freeze until i managed to close the tab, which took minutes, or the renderer process got killed by systemd-oomd, which only worked with pressure stall information enabled in the kernel because the GPU wedged so badly it would take hours to hit an OOM condition on the 512MB that were assigned to it
1
0
1
2
0
5
johnny peligro
mischievoustomato@0.5dollah.click@mia that's super odd. I wonder how stuff is because I routinely play games and often everything is fine. I only rarely hit a "oops gpu reset due to hang", but I havent hit a VRAM issue like this. Do you have any webpage you think that reproduces this? The only issue with a webpage I have is X.com opening an audiocontext thing that never suspends so it consumes cpu because of EasyEffects. Vibecoded a tampemonkey script to deal with that 😔 In general GPU stuff in linux still kinda feels messy.
1
0
0
@mia back in 2021? 2022? i had to get lucky to beat a boss in GI because eventually the gpu would reset.
1
0
0
johnny peligro
mischievoustomato@0.5dollah.click
0
0
0
@mischievoustomato yeah i had exactly one (1) GPU hang with games on amdgpu (during some cutscene in doom eternal), and it recovered from that just fine
2
0
1
johnny peligro
mischievoustomato@0.5dollah.click@mia when i had a ryzen laptop a successful recovery was about 1/10 of the time :(
1
0
1
@mischievoustomato one of the websites that triggered this is openweathermap.com but it doesn’t happen reliably on any website
it has a webgl map widget that insta kills the haswell iGPU
0
0
1
@mischievoustomato i managed to hang the AMD GPU with some heavy OCR task involving a 24GB multimodal LLM and it triggered a full GPU reset that killed niri because it couldn’t preempt the ML task to respond to niri’s commands in time. i changed the timeout from 2 to 4 seconds since it’s just the GPU being slow, and it never happened again. no idea what cursed kernels pytorch is running there but 99% of tasks don’t cause GPU stalling like that
1
0
1
johnny peligro
mischievoustomato@0.5dollah.click@mia yeah same. sometimes a game does it but it seems rare, i should tweak it for my igpu too. I'm on an alder lake one. it chokes when context is big enough on an llm or when doing inpainting or img2img on imggen with koboldcpp
0
0
0