One GPU, several VMs. Split it with SR-IOV vGPU — pick your GeForce card and how many VMs want it, and see how much VRAM each gets.
Not listed? Custom:
How this works. NVIDIA's vGPU (SR-IOV) splits a physical GPU into fixed-size profiles, each 1 GB granularity (the "Q" series — 1Q = 1 GB, 8Q = 8 GB, and so on). Each VM gets one virtual function (VF). The combined VRAM of all running vGPUs must fit in physical VRAM, with headroom left for the host driver. Consumer cards are presented as their professional sibling (Ampere → RTX A6000, Ada → RTX 6000 Ada), so the profile table can technically offer sizes larger than your card — we cap to physical VRAM here. This is an estimate: real limits depend on driver version and your host's IOMMU/BAR setup.