KubeInfer

MIG vs time-slicing: partitioning GPUs for inference

When to slice an A100 into seven, and when not to.

CR
Carlos Ruiz
Jun 26, 2026·1 min read·225

A single A100 is overkill for a 1B parameter model. The question is how to share it safely between workloads.

Multi-Instance GPU (MIG)

MIG gives you hardware-isolated partitions with dedicated memory and compute. Predictable, but rigid — profiles are fixed sizes.

Time-slicing

Time-slicing oversubscribes a GPU by context-switching between pods. Higher density, zero isolation. One noisy neighbor can wreck your p99.

  • Hard multi-tenant boundary? Use MIG.

  • Bursty internal workloads that tolerate jitter? Time-slice.

  • Latency-critical single tenant? Don't partition at all.

nvidia-smi mig -cgi 19,19,19 -C   # three 1g.10gb instances
CR

Written by

Carlos Ruiz

GPU performance nerd

Squeezing every last token/sec out of accelerators. CUDA, MIG, and kernels.

3 followers · 2 stories