GPU-aware Dell fan control for Proxmox

This repo contains a small two-part setup for using GPU temperatures from a passthrough VM/render node to control Dell PowerEdge chassis fans from the Proxmox host.

It is useful when:

• GPUs are passed through to a VM.
• iDRAC cannot see the real GPU temperatures.
• Dell's automatic fan logic does not react to GPU heat.
• Static manual fan speed is either too loud or not safe enough.

Architecture

GPU VM / render node
  nvidia-smi
      ↓
  /mnt/nas_ai/shared/gpu_fan_bridge/gpu_temp
      ↓
Proxmox host
  gpu-fan-control.sh
      ↓
  ipmitool raw Dell fan control

The VM exports:

/mnt/nas_ai/shared/gpu_fan_bridge/gpu_temp
/mnt/nas_ai/shared/gpu_fan_bridge/gpu_temps_detail
/mnt/nas_ai/shared/gpu_fan_bridge/heartbeat
/mnt/nas_ai/shared/gpu_fan_bridge/status

The Proxmox host reads those files and sets Dell chassis fan speed.

Files

scripts/gpu-temp-export.sh         # run inside the GPU VM/render node
scripts/gpu-fan-control.sh         # run on the Proxmox host
systemd/gpu-temp-export.service    # VM service
systemd/gpu-fan-control.service    # Proxmox oneshot service
systemd/gpu-fan-control.timer      # Proxmox timer

Requirements

GPU VM / render node

• NVIDIA driver working
• nvidia-smi available
• shared NAS/NFS mount available at /mnt/nas_ai/shared

Proxmox host

• ipmitool
• same shared NAS/NFS mount available at /mnt/nas_ai/shared
• Dell server that supports the Dell raw fan commands

Install ipmitool on Proxmox:

apt update
apt install -y ipmitool

Install on the GPU VM / render node

Copy:

sudo cp scripts/gpu-temp-export.sh /usr/local/bin/gpu-temp-export.sh
sudo chmod +x /usr/local/bin/gpu-temp-export.sh
sudo cp systemd/gpu-temp-export.service /etc/systemd/system/gpu-temp-export.service
sudo systemctl daemon-reload
sudo systemctl enable --now gpu-temp-export.service

Watch logs:

sudo journalctl -u gpu-temp-export.service -f

Check output:

cat /mnt/nas_ai/shared/gpu_fan_bridge/gpu_temp
cat /mnt/nas_ai/shared/gpu_fan_bridge/gpu_temps_detail
cat /mnt/nas_ai/shared/gpu_fan_bridge/heartbeat
cat /mnt/nas_ai/shared/gpu_fan_bridge/status

Install on the Proxmox host

Copy:

cp scripts/gpu-fan-control.sh /usr/local/sbin/gpu-fan-control.sh
chmod +x /usr/local/sbin/gpu-fan-control.sh
cp systemd/gpu-fan-control.service /etc/systemd/system/gpu-fan-control.service
cp systemd/gpu-fan-control.timer /etc/systemd/system/gpu-fan-control.timer
systemctl daemon-reload
systemctl enable --now gpu-fan-control.timer

Watch logs:

tail -f /var/log/gpu-fan-control.log

Check state:

cat /run/gpu-fan-state

Default fan curve

The Proxmox script uses fan levels rather than raw one-shot thresholds.

Temp	Level
<65C	32%
65C+	40%
72C+	50%
78C+	60%
84C+	80%
88C+	100%
92C+	immediate 100%

It also has:

• hysteresis
• minimum hold time
• one-step fan transitions
• stale heartbeat handling
• 100% failsafe for long stale data

Emergency quiet command

On Proxmox:

systemctl stop gpu-fan-control.timer
ipmitool raw 0x30 0x30 0x01 0x00
ipmitool raw 0x30 0x30 0x02 0xff 0x28

That sets manual fan mode around 40%.

Return to Dell automatic mode

ipmitool raw 0x30 0x30 0x01 0x01

Important safety note

This setup intentionally does not fall back to iDRAC automatic fan control when GPU data is stale. On systems where iDRAC does not see passed-through GPU temperatures, automatic fan control may be too slow or unaware of the GPU heat. Long stale data forces 100% instead.