Fixing RTX Pro 6000 Blackwell shutdowns with custom fan control

My dual RTX Pro 6000 Blackwell workstation would hard power off under sustained inference load. The fix ended up being simple: take control of the fan curve.

Hardware

CPU: AMD Ryzen 9 7950X3D 16-Core Processor
Motherboard: ROG CROSSHAIR X670E HERO
GPU: Dual NVIDIA RTX Pro 6000 (96 GB VRAM each)
RAM: 192 GB DDR5 5200

Symptom

During long-running jobs the machine would abruptly lose power (no graceful shutdown). The default fan curve is tuned for low noise and is slow to ramp under sustained high power draw.

What I changed

I run a small Python daemon that uses NVML (NVIDIA Management Library) to set fan speed based on temperature:

• Poll temperature once per second
• Apply an aggressive curve that still idles quietly
• Run as a systemd service so it starts on boot

The curve I’m using:

Temperature	Fan speed
≤40°C	30%
50°C	55%
55°C	75%
60°C	90%
≥65°C	100%

Once the GPUs approach 60–65°C the fans are already high, which keeps them away from the thermal cliff.

Install

Tested on Ubuntu 24.04.

git clone https://github.com/zmarty/nvidia-fan-control
cd nvidia-fan-control

sudo apt install python3-pynvml

sudo mkdir -p /opt/nvidia-fan-control
sudo cp nvidia-fan-control.py /opt/nvidia-fan-control/
sudo chmod +x /opt/nvidia-fan-control/nvidia-fan-control.py

sudo cp nvidia-fan-control.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable nvidia-fan-control.service
sudo systemctl start nvidia-fan-control.service

Logs:

journalctl -u nvidia-fan-control -f

Code: github.com/zmarty/nvidia-fan-control

Validation

I load-tested with a local vLLM server and a small client that sends 50 concurrent requests in a loop. With the custom curve in place, the workstation stays up under sustained load.