Hardware: Jetson Thor T5000
Software:
-
nvidia-jetpack: 7.2-b187
-
nv_tegra_release: R39 (release), REVISION: 2.0
I enabled MIG by following the official documentation and created two GPU Instances (GIs):
$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
MIG 2g.0gb Device 0: (UUID: MIG-64c6ac3f-05fa-5ae0-8c2e-ec6fb3d6f79e)
MIG 1g.0gb Device 1: (UUID: MIG-e08de158-303a-5b82-b48c-49d023899e20)
I exported CUDA_VISIBLE_DEVICES in two separate terminals:
export CUDA_VISIBLE_DEVICES=MIG-64c6ac3f-05fa-5ae0-8c2e-ec6fb3d6f79e
export CUDA_VISIBLE_DEVICES=MIG-e08de158-303a-5b82-b48c-49d023899e20
CUDA applications work correctly on both MIG instances. For example:
1. CV-CUDA Resize
| Profile | Mean Time (ms) |
|---|---|
| 2g.0gb (+gfx) | 0.0733 |
| 1g.0gb (+me) | 0.1214 |
The performance difference is roughly consistent with the expected compute resource ratio.
2. VIC
The application hangs when using the VIC backend.
3. VPI CUDA Resize
| Profile | Mean Time (ms) |
|---|---|
| 2g.0gb (+gfx) | 0.0909 |
| 1g.0gb (+me) | 0.1099 |
4. OFA (02-stereo_disparity)
2g.0gb (+gfx):
line 203: VPI_ERROR_INTERNAL: (:NV_OF_ERR_GENERIC)
1g.0gb (+me):
line 207: VPI_ERROR_INTERNAL: VIC|OFA backend(s) error: (:NvSciError_BadParameter)
5. PVA (01-convolve_2d)
2g.0gb (+gfx):
VPI_ERROR_INTERNAL: VIC|OFA backend(s) error: (:NvSciError_BadParameter)
1g.0gb (+me):
VPI_ERROR_INTERNAL: VIC|OFA backend(s) error: (:NvSciError_BadParameter)
It appears that none of the SoC engines (VIC, PVA, or OFA) work after MIG is enabled.
I’d like to know whether I’m using MIG incorrectly, or if the current MIG implementation simply does not support these SoC engines (VIC, PVA, and OFA). If they are supported, are there any additional configuration steps required?