Infrastructure · Infrastructure resource

Time-Sliced GPU Share

Infrastructure resourceInfrastructureInfrastructurearc:TimeSlicedGPUShare

A share of a whole GPU in which the driver schedules co-located processes in round-robin time slices over a single shared memory pool, without hardware isolation.

Responsibility. Lets several workloads share one GPU flexibly, including bursting into memory left idle by others.

Also known as: GPU time-slicing, Software GPU sharing

Variant of GPU Allocation Unit abstract

When to choose. Choose for internal development, trusted co-located workloads, long-running batch jobs, or cost-constrained deployments that tolerate latency variance and cannot afford partitioning overhead.

Direct neighbourhood (hover for relationship types)

Relationships

hosts structural

Inference Server Ch7.6

alternative to variability

Design guidance

SHOULD NOT host SLA-critical, multi-tenant or latency-sensitive agent inference.

Quantitative guidance

As stated by the sources; verify before use.

Time slices are typically 50-200 ms; a long co-tenant kernel can block others for 500 ms+ (Ch7.6).
4 processes on A100-80GB (7B, batch 1): P50 120 ms, P99 450 ms, +/-250 ms vs MIG 1g.10gb P50 75 ms, P99 95 ms, +/-15 ms (Ch7.6).

Classification

Technologies: CUDA driver time-slicingNVIDIA vGPU
Quality attributes: Flexibility (ISO/IEC 25010)Cost efficiency

Sources

Ch7.6: T. Nguyen, "Multi-Instance GPU," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 7.6. ISBN: 9798244538229.