08/13/2026
Scheduling a GPU fraction ≠ enforcing a GPU fraction. Your scheduler can allocate 0.25 of a GPU to a pod. That’s an accounting decision — it says nothing about what stops that pod from using more once it’s actually running.
Think of it like a kitchen with multiple cooks sharing the same equipment:
🔥Time-slicing: cooks take turns at one stove. Nothing stops a cook from hogging every burner. No reservation, no enforcement, nothing keeps a workload inside its slice.
🎛️ MPS: a kitchen manager coordinates who cooks when. A control daemon enforces memory/compute division in software: effective, but it’s a single point of failure for everything on the GPU.
🧱 MIG: fully separate stations, walled off from each other. Hardware-partitioned, fully isolated. Strongest guarantee, but station sizes are fixed until you reconfigure the whole kitchen.
If you’re only looking at the scheduler, you’re missing the layer that decides whether “25%” actually means anything at runtime.
Julien breaks this down, plus how to pair schedulers like KAI with enforcement layers like HAMi-core, in our latest webinar, visit 🔗 kubex.ai/resources/sharing-gpus
