Dell Poweredge R720
Nvidia Tesla P40 24GB
GPU pass-through via Proxmox
-
? am using only one 1100w psu, the other 1100w is not plugged in.
-
dont have a UPS yet.
- upgrade idrac firmware, so you get to power on/off via http.
- if installing at riser3, ensure the pci-e is x16 - some servers may come with a riser with 2 slots each with x8.
- Please be mindful when purchasing the gpu power cable, there are 2 kinds of gpu power cables for dell servers PCI risers, one for the Nvidia Telsas and one for consumer "general purpose" GPUs. This is a step that SHALL NOT GO WRONG or else you may fry your server&GPU! read more here: https://kenmoini.com/post/2021/03/fun-with-servers-and-gpus
- determine if you should enable/disable "3rd Party Card fan behavior"
- (?) enabling this allowed the GPU temp to hover at around 60C - while being used, Fan hovers in the range of 4200 to 7000 RPM. (typical GPU temp in a server room with AC will be around 55C under load). The Inlet and Exhaust Temp at 31 and 38C respectively (server is placed in the garage with no AC, with outdoor temp at 29C).
http://www.righteoushack.net/dell-poweredge-13th-gen-fan-noise/ https://www.reddit.com/r/Proxmox/comments/uf2d7l/proxmox_tesla_m40_passthrough_ubuntu_server_vm/iif2en3/?context=3
- current settings, likely not optimal.
ssh idrac
racadm set system.thermalsettings.AirExhaustTemp 255
racadm set system.thermalsettings.FanSpeedOffset 0
racadm set system.thermalsettings.ThermalProfile 0
racadm set system.thermalsettings.ThirdPartyPCIFanResponse 1
racadm get system.thermalsettings
Fan at 9480RPM
GPU Temp 60 C per nvidia-smi
Fan1 | 8520 RPM | ok
Fan2 | 8400 RPM | ok
Fan3 | 8520 RPM | ok
Fan4 | 9360 RPM | ok
Fan5 | 10560 RPM | ok
Fan6 | 9960 RPM | ok
Inlet Temp | 33 degrees C | ok
Exhaust Temp | 37 degrees C | ok
Temp | 55 degrees C | ok
Temp | 46 degrees C | ok
Current 1 | 2.60 Amps | ok
Current 2 | no reading | ns
Voltage 1 | 110 Volts | ok
Voltage 2 | no reading | ns
Pwr Consumption | 294 Watts | ok
Chassis Temp 38C
-
follow below link and stop prior section "Configuring the VM (Windows 10)", note the modifications listed below. https://gist.github.com/baodrate/64f617e959725e934992b080e677656f
-
in proxmox web interface, select vm, for hostpci - check
All Functions,ROM-Bar,PCI-Express -
for vm BIOS, use
Default (SeaBIOS). -
for vm Machine, use
q35. -
edit vm conf
/etc/pve/qemu-server/${VM_ID}.confper below
-
cpu: host,hidden=1,flags=+pcid
args: -cpu 'host,+kvm_pv_unhalt,+kvm_pv_eoi,hv_vendor_id=NV43FIX,kvm=off'
# below is based on needs
cores: 10
memory: 262144
scsi0: local-lvm:vm-100-disk-0,size=768G
boot: order=scsi0;net0
scsihw: virtio-scsi-pci
hostpci0: 0000:42:00,pcie=1
-
turn on the vm.
-
when installing ubuntu, don't install the driver.
-
boot up vm
-
check if gpu is present
lspci | grep 01:00
- install driver
sudo apt-add-repository -r ppa:graphics-drivers/ppa
sudo apt update
sudo apt remove nvidia*
sudo apt autoremove
sudo ubuntu-drivers autoinstall
- run
nvidia-smiand get complaints, run below.
sudo rmmod nouveau
sudo modprobe nvidia
# https://unix.stackexchange.com/questions/219059/remove-nouveau-driver-nvidia-without-rebooting
-
run
nvidia-smito confirm presense of gpu. -
install nvidia container toolkit https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html
-
confirm install success
docker run --runtime=nvidia --rm nvidia/cuda nvidia-smi
That's weird, before I got the P40 I was using the K80. For the K80 I got a different issue, I use a single 8-pin to 8-pin cable to connect RISER and the GPU. Then in Windows 11, it only shows 1 GPU is working, the second GPU shows code 12 too.