In theory yes, and we got tiny programs running this way, but could not afford to keep subsidizing megacorp politics.
Nvidia punts to IBM RHEL8 docs for GPU podman, which is unusual and risky to see. We officially recommend against it for HA environments due to this kind of lack and overall low relative confidence. I think k8s envs may be moving to something here, so maybe in a year or two? I'd be curious of folks doing stock rhel8 podman with tensorflow/torch on nvidia, which should be as vanilla as you can get for enterprise ai. We generally see more interesting GPU envs here (ex: DGX with advanced networking hw/sw), but we don't have confidence for the simple case, which is the starting point..
Nvidia punts to IBM RHEL8 docs for GPU podman, which is unusual and risky to see. We officially recommend against it for HA environments due to this kind of lack and overall low relative confidence. I think k8s envs may be moving to something here, so maybe in a year or two? I'd be curious of folks doing stock rhel8 podman with tensorflow/torch on nvidia, which should be as vanilla as you can get for enterprise ai. We generally see more interesting GPU envs here (ex: DGX with advanced networking hw/sw), but we don't have confidence for the simple case, which is the starting point..