Multi-Instance GPU
An Nvidia feature that splits one data centre GPU into as many as seven smaller, isolated GPUs.
Also: MIG, GPU partitioning
Introduced with the A100 in 2020, Multi-Instance GPU divides a single GPU's cores, memory and cache into up to seven separate instances, each with guaranteed resources and isolated from the others. It lets a provider rent a fraction of an expensive GPU for small models, notebooks or light inference. Other methods, such as time-slicing, share a GPU by taking turns but without hard isolation. Fractional GPUs help keep utilisation high when many users need less than a full chip.