How to make the aodt backend to use 2 GPU's

Accelerated Computing Aerial Forum
,
1

Hello Team,

I am working on the aodt v1.2 and right now I am able to work on a backend GPU of A100. I would like to know what setting do i need to tweak so that I can use both of the GPU’s in the A100 server card.

My setting currently is
env:
- name: AODT_SIM_GPU
value: ‘1’
- name: NVIDIA_VISIBLE_DEVICES
value: ‘1’
- name: NVIDIA_DRIVER_CAPABILITIES
value: ‘all’

i tried setting the visible devices to ‘0,1’. The pod then shows both GPU’s but i can clearly see that 1 GPU is free and not used.

aerial@nucleus-omni-worker-96dbdc996-z7ndd:/aodt/aodt_sim/build$ nvidia-smi
Mon Mar 3 05:11:28 2025
±----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.90.07 Driver Version: 550.90.07 CUDA Version: 12.6 |
|-----------------------------------------±-----------------------±---------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA A100 80GB PCIe On | 00000000:DA:00.0 Off | 0 |
| N/A 53C P0 86W / 300W | 919MiB / 81920MiB | 0% Default |
| | | Disabled |
±----------------------------------------±-----------------------±---------------------+
| 1 NVIDIA A100 80GB PCIe On | 00000000:DB:00.0 Off | 0 |
| N/A 41C P0 43W / 300W | 4MiB / 81920MiB | 0% Default |
| | | Disabled |
±----------------------------------------±-----------------------±---------------------+

Thanks,
Sujith.

2

@sujith.samuel Multi- GPU is not supported yet. The backend will run on a single GPU only.

3

Hello @kpasad1 is this feature in the pipeline???

might be good to have for bigger simulations. for a 6UE,1DU,1RU sim i am seeing backend take up 32GB GPU.

Thanks,
Sujith.

4

@sujith.samuel
The feature is in pipeline. 32gb is not unreasonable. Looks like you have enabled RAN. When RAN is enabled, the number of UEs scheduled per RU is limited, so even if you scale it up to a larger number of UEs, the memory consumption will not increase linearly beyond a certain point. A PF scheduler will ensure that all UEs are scehduled.