An operations team is choosing between two NVIDIA vGPU profile families on a data-centre GPU. One family backs each vGPU with a dedicated hardware partition that has its own memory path and compute engines, while the other lets multiple vGPUs share the whole GPU through scheduled time slices. They want the family that delivers spatial isolation with guaranteed compute throughput per instance rather than scheduled access to the full device. Which underlying GPU capability provides that spatial, hardware-partitioned isolation?
- AMulti-Instance GPU partitioning, which divides the GPU into hardware-isolated instances each with dedicated compute slices, memory, and cache, so a vGPU mapped to an instance gets guaranteed throughput. Correct
- BRound-robin time-slicing of the full GPU, which rotates each vGPU through the complete set of compute engines for short intervals to approximate dedicated access for every instance.
- CSingle Root I/O Virtualisation, which exposes the GPU as multiple virtual functions on the PCIe bus so the hypervisor can map one virtual function per guest.
- DUnified memory addressing, which lets each vGPU oversubscribe physical memory by migrating pages between host and device on demand during execution.
Why A is correct: Correct: MIG physically partitions streaming multiprocessors, L2 cache, and memory controllers into separate instances, so a MIG-backed vGPU receives spatially isolated, guaranteed compute rather than a scheduled share of the whole GPU.
Why B is wrong: Tempting because it is a real vGPU sharing mode, but time-slicing is temporal not spatial: instances take turns on the whole device, so throughput is not isolated and a noisy neighbour can still affect timing, failing the spatial-isolation requirement.
Why C is wrong: Tempting because SR-IOV is part of how MIG-backed vGPUs are presented to guests, but SR-IOV is only the PCIe presentation mechanism; on its own it does not partition compute and memory into hardware-isolated slices.
Why D is wrong: Tempting because unified memory is a genuine CUDA feature, but it governs page migration and oversubscription, not partitioning the GPU into isolated instances, so it provides neither spatial isolation nor guaranteed per-instance throughput.