How close in perfromance is the native timeslicing using NVIDIA GPU Operator compared to using vGPU (vWS) for both graphics and compute workloads? Secondly which has better (lower) latency of the two during context switches for graphics?
What makes me ask this question is, does the absense of a hypervisor reduce overhead when someone uses time-slicing in bare metal K8s with GPU Operator, therefore lower latency or does vGPU (via vGPU Manager) more efficiently handle the context switching compared to native time-slicing offered by GPU Operator?
I’ll appreciate your perspective/help. Thanks.