Multi-Instance GPU
GPU Hardware & ComputeA capability that partitions a single physical GPU into several fully isolated instances, each with its own dedicated memory and compute, so one chip can serve multiple workloads at once.
Multi-Instance GPU is the technology that makes fine-grained, flexible GPU access possible, complementing the dedicated, bare-metal end of Eleveight AI's range with the ability to right-size compute to smaller workloads.
Overview
Not every workload needs a whole GPU. A modern accelerator is so capable that a small inference service or a modest experiment may use only a fraction of it, leaving the rest idle. Multi-Instance GPU, or MIG, addresses this by splitting one physical GPU into several smaller, independent instances, each behaving like a self-contained GPU with its own slice of memory and processing cores.
How it works
MIG partitions the GPU at the hardware level, carving its memory and compute resources into separate instances that are genuinely isolated from one another. A workload running in one instance cannot affect or observe the others, and each receives a guaranteed, dedicated share rather than competing for a common pool. The partitioning can be configured to suit the mix of workloads a chip needs to host.
Why it matters
Hardware partitioning improves utilization and makes compute divisible to the size of the job. Rather than handing a whole expensive GPU to a task that needs a tenth of it, an operator can serve many smaller workloads from one chip, each with firm isolation and predictable performance. This is what allows flexible, scalable access to be offered economically alongside whole-GPU and whole-cluster reservations.
Use cases
- Serving multiple small inference workloads on one GPU
- Right-sizing compute for modest jobs
- Improving utilization in shared environments
- Isolating tenants on partitioned hardware