MIG is an NVIDIA hardware feature that physically partitions a supported GPU into up to seven fully isolated instances, each with its own memory and compute cores.
Unlike software-level vGPU sharing, MIG isolation happens at the silicon level, which makes it well suited to strict multi-tenant SLAs where noisy-neighbor performance issues aren't acceptable.
hosted·ai's scheduler works alongside hardware-level isolation like MIG as part of its multi-tenant scheduling model. See how hosted·ai schedules multi-tenant GPUaaS workloads.
We use cookies for analytics and advertising. Privacy policy