Previously, nvidia hid hotspot monitoring from consumers because fuck you
Recently some Brazilians got a hold of diagnostic software and extracted what to look for.
Turns out the hotspot in some conditions can be 30°C higher than what the card will normally report, which is why you can end up with unexpected throttling.
Hotspot reporting for these cards is now starting to show up in software like HWMonitor
Is this the first for any GPU company? I wonder if NVIDIA’s shenanigans will ever lead to any meaningful gain in market share for their competitors, but it looks like with AMD and Intel also going all in on AI they’ll pull something too (or intel will just kill arc)
Should note, that if someone has figured out how to pull out better hotspot data from an NVidia card, someone else could fold that into the (open-source) Linux driver for NVIdia cards. You might not even need to be especially acquainted with NVidia or Linux drivers, if you enlist an LLM to help.
In a recent exercise I asked an LLM to introspect current and older versions of the AMD video driver(s), and the LLM did surprisingly well. I was building a fan-control service for an old AMD datacenter GPU that lacked working fan-control. The LLM was able to ensure that the service was entirely proper and that could be expected to work with different driver versions. The LLM was even able to spot that the “common knowledge” names for the different hotspot measures were slightly wrong.
I do not have a recent NVidia card, so lack interest and means of test.
