README.md

vramtemp

VRAM Temperature Reader

Ampere and Ada don't report VRAM temperatures in official NVIDIA Linux tools such as nvidia-smi / nvidia-settings. So we work around it by reading directly from the driver's outputs, with a dead simple C program, easy to consume by other scripts and applications. Tested on RTX 3090.

Note

This is a barebone rewrite of gddr6 by Ole Algoritme (@olealgoritme).
The purpose of vramtemp is compliance with stdio: stdout (>1) is a single reading then exits.

• Works with pipes | > && etc.
• Consumable by scripts.
• Non-blocking for the terminal session. Pipe with watch or log to monitor continuously.

Core logic stays the same.

Build

One file: vramtemp.c

gcc -o vramtemp vramtemp.c -lpci

Usage

By default, outputs values in one line, space-separated.

sudo ./vramtemp

↓

76 42

Flag -l to get one GPU per line.

sudo ./vramtemp -l

↓

0 76
1 42

Use awk or cut to get your desired value.

sudo vramtemp | awk '{print $1}'   #→ 76
sudo vramtemp | awk '{print $2}'   #→ 42
sudo vramtemp | cut -d' ' -f1      #→ 76
sudo vramtemp | cut -d' ' -f2      #→ 42

Install

Just copy the executable to a dir in your PATH.
/usr/local/bin fits the FHS.

Why vramtemp?

• Fast (written in C).
• Minimal dependencies: NVML is not required (no nvml.h, no CUDA).
• Clean output then exit. Use e.g. watch -n 0.5 or similar tool to get a continuous read.
• Tool-friendly: works with scripts, pipes, etc.
• Extremely small: less than 2,000 chars (1.8KB).

Thanks

The heavy lifting was entirely done by Ole Algoritme (@olealgoritme) in gddr6, reverse-engineering the NVIDIA driver!
I only trimmed it for my use case (with Grok 4.3 (beta)).