I’ve recently acquired two RTX 6000 Ada Lovelace GPUs to replace my 4090 rig and am noticing out of the box that the RTX 6000 Adas are running a lot hotter than the 4090s did.
I may not be used to professional cards, but I did notice that the GPU fan speed does not reach 100% when temperatures start exceeding 80c like my 4090 does and that on Linux I have no easy way to configure the fan curve so that it does.
The GPUs are racked and have high airflow fans pushing air through the case, so noise is not a concern for me.
This in your case reads to: you run at 88°, which is 4° away from SW slowdown, which for this sku would then be 92°. HW slowdown would be -2°, so 2° above that 92°, with a HW SHUTdown kicking in at -7°. We no longer show you fixed values (that would be different per GPU), but only your delta to that, and your absolute average temp at any time…
This blows my mind. Since atleast Ada NVML has been reporting these values and no one at Nvidia has explained why exactly it shows those values. nvidia-settings started showing correct looking values, but it was never clear where they came from. I looked through the code but either missed it or it’s somewhere else. No one could spend a little time to update the header? Or just do a new API since it’s so radically different than the old one?
It’s not even a very great way of doing it. The temp it uses as the base temp is the max temperature limit you can set via “acoustic threshold”, but on some GPUs(like mine), that isn’t supported, so it’s the result of:
base_temp = current_temp + margin_temp
Not only did the margin temp API not even exist until well after Ada was released, it’s not even very reliable because you’re polling two different values at different times. Your base_temp could be 87 but depending on when you poll it in a reasonable time frame, it could be 86, 87, or 88.
Can Nvidia please stop adding broken APIs and give people access to your raw APIs? No third party developer can read your minds and handle how broken this is and you clearly have no intention of fixing your broken and poorly documented higher level ones.
I decided to double check if it was or wasn’t supported since I was assuming that since no current value could be read(NVML_TEMPERATURE_THRESHOLD_ACOUSTIC_CURR) that the min/max values were also unsupported but no, on a desktop RTX 5060 TI 16GB, it is not possible to set a temperature limit or get the upper or lower bound temperature:
