Enabling GPU support

This guide outlines how to enable GPU support on your CRN, as to allow user do deploy Instance with GPUs.

Enabling GPU support for Virtualization requires changing which Kernel module (driver) is handling the GPU card, as it need the special vfio module to works with qemu. Do note that it will make the GPU unavailable for other purposes.

To activate this feature it is also required to enable PAYG support and IPv6 on your CRN, as this is the required payment method for GPU. Follow the steps at Enable PAYG

Device compatible with the feature

At the time of writing the GPUs listed below are compatible with the feature, more will be added soon as they are tested and validated.

Standard GPUs

GPU Model	vCPU	RAM	vRAM	Price approx
L40S	12	72 GB	48 GB	3.33 ALEPH/hour
RTX 5090	8	48 GB	36 GB	2.24 ALEPH/hour
RTX 4090	6	36 GB	24 GB	1.68 ALEPH/hour
RTX 3090	4	24 GB	24 GB	1.12 ALEPH/hour
RTX 4000 ADA	3	18 GB	20 GB	0.84 ALEPH/hour

GPUs must be connected via PCIe 4.0 16x each.

Premium GPUs

Datacenter grade GPUs.

GPU Model	vCPU	RAM	vRAM	Price approx
H100	24	144 GB	80 GB	13.33 ALEPH/hour
A100	16	96 GB	80 GB	8.89 ALEPH/hour

Switch Kernel modules for GPU

It is possible to enable multiple GPUs on one CRN

Execute these command as root, use sudo on Ubuntu system.

Edit initramfs to attach GPU to vfio: vi /etc/initramfs-tools/modules and set the content to
```
attach : vfio vfio_iommu_type1 vfio_virqfd vfio_pci ids=10de:27b0,10de:22bc
```

replacing the ids 10de:27b0,10de:22bc with ids of the VGA card, there should be 2, the Video device and the GPU audio device. You can get them running lspci -nvv as root user.

Edit /etc/modules to ensure that lods GPU to vfio: vi /etc/modules and add
```
attach: vfio vfio_iommu_type1 vfio_pci ids=10de:27b0,10de:22bc
```
Modify nvidia drivers to load after the vfio: vi /etc/modprobe.d/nvidia.conf

set

softdep nouveau pre: vfio-pci
softdep nvidia pre: vfio-pci
softdep nvidia* pre: vfio-pci

Get the know alias from the PCI device: cat /sys/bus/pci/devices/0000:01:00.0/modalias replace pci address 0000:01:00.0 with the one for your GPU, you can get it using lspci
Configure vfio module to use that devices: vi /etc/modprobe.d/vfio.conf

blacklist nouveau
blacklist snd_hda_intel
alias pci:v000010DEd000027B0sv000010DEsd000016FAbc03sc00i00 vfio-pci
options vfio-pci ids=10de:27b0,10de:22bc

Enable modprobe vfio-pci
```
modprobe vfio-pci
```

7 . Update initramfs image with your changes:

update-initramfs -u -k all

Finally Reboot the server
```
reboot
```
Confirm that the vfio-pci kernel module is used for the card
```
lspci -k
```

Should display something similar to

[...]
01:00.0 VGA compatible controller: NVIDIA Corporation AD104GL [RTX 4000 SFF Ada Generation] (rev a1)
    Subsystem: NVIDIA Corporation AD104GL [RTX 4000 SFF Ada Generation]
    Kernel driver in use: vfio-pci
    Kernel modules: nvidiafb, nouveau, nvidia_drm, nvidia
01:00.1 Audio device: NVIDIA Corporation Device 22bc (rev a1)
    Subsystem: NVIDIA Corporation Device 16fa
    Kernel driver in use: vfio-pci
    Kernel modules: snd_hda_intel
[...]

Enable GPU support in Aleph-VM In the CRN configuration /etc/aleph-vm/supervisor.env, enable the GPU support
```
ALEPH_VM_ENABLE_GPU_SUPPORT=True
```

Don't forget to also set ALEPH_VM_PAYMENT_RECEIVER_ADDRESS and enable PAYG.

Confirm that the GPU are listed and supported on the CRN index page.

Start or restart your aleph-vm supervisor, open the index page, and it should list your GPU

GPUs

   • NVIDIA | AD104GL [RTX 4000 SFF Ada Generation] is compatible ✅

Disabling support.

To disable support for the GPU, revert the modification done to attach the kernel module to the GPU.