Multi-Instance GPU (MIG)

Multi-Instance GPU, or MIG, is an NVIDIA feature that partitions a supported GPU into as many as seven isolated instances, each with dedicated memory, cache, and compute cores. Each instance behaves like a standalone GPU to applications, so CUDA code runs unchanged. A failure in one instance does not affect the others. Dedicated memory bandwidth per instance also stops a busy neighbor from starving a job.

Profile names encode the slice. On an 80 GB H100, 1g.10gb is one seventh of the streaming multiprocessors and one eighth of the memory, and seven fit on the card. A 7g.80gb profile takes the whole GPU. The A100, H100, H200, B200, and B300 allow seven instances. The RTX PRO 6000 Blackwell, in Server, Workstation, and Max-Q editions, allows four, as 1g.24gb slices of 24 GB each. The A30 allows four, and the RTX PRO 5000 and 4500 Blackwell allow two. Unlike time-slicing, MIG runs each workload in parallel on its own hardware.

Use MIG when several tenants or inference services must share one GPU with guaranteed memory and predictable latency. Then size the card by slice count and memory. An 80 GB H100 yields seven 10 GB slices, and an RTX PRO 6000 Blackwell yields four 24 GB slices.

Sources

Last verified October 7, 2026.