Virtualization

Technology that allows creating multiple virtual instances (VMs) from a single physical machine, enabling better resource utilization and isolation. In GPU computing, virtualization can be challenging because GPUs traditionally required direct hardware access. Modern solutions include GPU passthrough (dedicating entire GPU to one VM), GPU partitioning (splitting one GPU among multiple VMs with NVIDIA MIG on A100/H100), and vGPU (NVIDIA software that time-slices GPU resources). Cloud providers use virtualization to offer flexible GPU instances, though some performance-sensitive workloads prefer bare-metal servers to avoid virtualization overhead.