Time-slicing provides no throughput isolation - one tenant's load silently degrades all others

One tenant runs a large batch job, and all other tenants on the same GPU experience latency degradation. There is no isolation.

Time-slicing (the default GPU sharing mode) schedules GPU kernels from different processes in a time-shared manner. There is no throughput guarantee - a process that submits large kernels can monopolize the GPU. Unlike MIG, time-slicing does not partition SMs or memory, so there is no isolation.

Use MIG for hardware-level isolation. Alternatively, use separate GPUs per tenant. For development/testing, time-slicing is acceptable; for production multi-tenant serving, MIG or separate GPUs are required.

Time-slicing is the default GPU sharing mode and provides zero isolation. It is suitable only for development and testing. For production multi-tenant serving, MIG is the minimum acceptable isolation level. The tradeoff is that MIG has its own limitations (no NCCL, shared NVLink).