Fractional GPUs 101: why slicing beats renting

The real problem: instance-sized walls

GPU demand is spiky. Instances are not. That mismatch is where budget dies.

What “fractional” actually means

Fractional compute is a scheduler promise: you get isolation + predictable performance, not a best-effort timeshare.

MIG vs MPS (and why you usually want both)

• MIG partitions hardware resources.
• MPS improves multi-process utilization.

What Kova optimizes for

Kova targets $/result, not $/hour.

A quick mental model

Slice policies define constraints → scheduling packs slices → receipts verify outcomes.