RTX 3080/3090 Memory/VRM temperature sensor

@Martin Hi martin, I am a linux user and currently training models using my 3090. Many linux users are pretty concern about the memory temps like me but we haven't found any linux software can readout this junction temp yet. In this post you mentioned that you get this value via nvidia driver and I'm pretty interested in how , just maybe I can try to readout the value by myself. Can you share more detailed information for us linux users? Thank you very much.
 
Last edited:
Has anybody noticed VRAM temps increasing over time?

I have modded my card with GeLid Extreme 1 mm on the VRAM and 3 mm GeLid Ultimate behind the vram areas of my rtx 3080.
I was observing 72°C for a good time, then I noticed 76°C, 80°C, and now I'm seeing 82°C appear regularly.
Before I never saw it pass 76°C peak temps, now I'm seeing 86°C peaks.
Ambient temperature hasn't changed enough to account for these swings.

Edit: looks like my "graphene" back plate started to bow, which may have loosened the cooler to the gpu/vram pressure. I opened it up to inspect it and nothing was out of ordinary. So I just reapplied paste and tossed it back in my system. I'll just keep using the 3mm on the back of the PCB because I don't really care about a small bit of bowing on the "graphene" back plate.
 
Last edited:
Maybe the pads are starting to loose their elasticity and reduces the ability to firmly touch the surfaces. I’m just speculating and don’t know for a fact.
 
I'm not planning a Linux version in near future and I have no idea whether nvidia-smi supports this either.
It would be much appreciated if you could throw a bone to concerned Linux users and even quickly describe how you get this info. We'd more than happily work on the implementation ourselves!

In any case, thank you for figuring this out, it is super useful for my 3080 even if it means I have to run windows sometimes :)
 
I think the way how it's read on Windows is quite Windows-specific and it's rather unlikely that it would work on Linux the same way.
 
GDDR6 ones don't seem to report memory temperature (internally in driver), so I'm afraid this probably won't be possible.
GPU-Z is able to report that on my 2080 Ti (GDDR6). So i think there is a way too read it on older cards too. I could inquire the developer how he did that and report back.
 
GPU-Z is able to report that on my 2080 Ti (GDDR6). So i think there is a way too read it on older cards too. I could inquire the developer how he did that and report back.

I checked that but I don't think GPU-Z is reporting the correct GDDR6 temperature. I verified it with a dedicated NVIDIA tool and the value GPU-Z reports belongs to a certain superset of temperatures, but I don't think it's the GDDR6.
 
@Martin Can you tell me the memory junction temp reported in the HWiNFo is for which memory chips on the pcb? or it just reports the average temp of all the chips or it reports the hottest chip only?
 
@Martin Can you tell me the memory junction temp reported in the HWiNFo is for which memory chips on the pcb? or it just reports the average temp of all the chips or it reports the hottest chip only?

It should be the hottest value among all memory chips.
 
  • Like
Reactions: gfx
Back
Top