Fractional GPU Allocation Across Enterprise AI Teams

Quick Answer: Fractional GPU allocation gives a team a stated share of one accelerator, plus a memory and isolation rule, so other teams can use the rest. Naming MIG or time-slicing is a mechanism choice. It is not an allocation policy.

Fractional GPU allocation is a scheduling policy that assigns part of a physical accelerator to a named team or job so a multi-team cluster can pack work without pretending every request needs a full card. The policy names the share, the isolation, and the reclaim. The driver feature only implements the share.

This page is the policy. Separate pages already compare MPS versus MIG and MIG versus time-slicing. Do not treat those mechanism pages as a quota design.

What must a fractional allocation name?

Write the five fields before you pick a mechanism. A guaranteed 20 GiB slice with hard isolation is a different product from a best-effort time slice. OnePlus Platform, OneSource Cloud's AI orchestration platform, can enforce team quota and show who holds a slice. Quota without reclaim is a reservation in disguise.

Which workloads belong on a fraction?

Interactive notebooks, small evals, token-light classifiers, and compile or tokenizer jobs often fit a fraction. Full-parameter training, large tensor-parallel ranks, and tight tail-latency serving usually do not. If the job’s minimum memory is most of the card, give it the card. A 0.9 fraction is a political full GPU.

Do not put two production chat replicas on time-sliced shares and call that high availability. Their tails move together. Prefer a full dedicated GPU, or a hard partition that actually isolates memory, when the SLO is a product promise.

Multi-team research clusters benefit first. Product inference benefits only when the share is hard and the neighbor class is boring. Private AI infrastructure still helps: exclusive nodes keep other companies off the card even when two internal teams share it.

How do you prevent one team from collecting fractions?

Cap concurrent fractions per team. Count memory, not only slice count, so ten 0.1 requests cannot hide a full card. Preempt or checkpoint interactive work that sits idle past the reclaim timer. Publish the hold time next to the quota, or teams will treat a fraction like a lease they never return.

Fair-share and guaranteed quota are different promises. Fair-share can shrink a slice when the cluster is busy. Guaranteed quota cannot. Do not label a fair-share slice as production capacity. The serving SLO page and the quota page will disagree.

U.S. exclusive halls, including Texas / Richardson capacity, do not change the arithmetic. They change who else might be on the PCI bus. Internal teams can still monopolize fractions. Managed AI infrastructure can run the reclaim. The policy owner still sets the timer.

What should you refuse?

Refuse a cluster that advertises fractional GPUs with no memory number. Refuse mixing PHI-class fine-tunes with student notebooks on a soft share. Refuse using average SM% as proof that sharing is safe. Refuse a mechanism bake-off that never writes the five fields.

FAQ

Is a fractional GPU the same as MIG?

No. MIG is one isolation mechanism that can implement a fraction. Time-slicing and some software schedulers implement other fractions. Choose the mechanism after the share, isolation, and class are written. Do not start from the acronym.

Can we train on a fractional GPU?

Only when the job’s memory and communicator fit the slice and you accept slower steps. Most multi-GPU training ranks want a full device. Use fractions for debug and small fine-tunes. Move the production train to whole cards.

Does fractional allocation replace team GPU quota?

No. Quota is how many GPUs, or GPU-equivalents, a team may hold. Fractional allocation is how a single device is carved. You need both. Otherwise a team with a small quota still pins an entire card at 0.05 used.

Should production inference use fractions?

Use them only with hard isolation and a neighbor class that cannot steal memory. Many enterprises keep production on full dedicated GPUs and leave fractions to notebooks and batch eval. That split is a policy, not a hardware limit.

How does OneSource Cloud relate to this policy?

Exclusive private GPUs reduce external noisy neighbors. OnePlus Platform can show and cap team shares. Neither writes the reclaim timer for you. You still name the five fields and refuse silent 0.9 “fractions.”

Summary

Fractional GPU allocation is a five-field policy: share, isolation, shape, reclaim, and class. Mechanisms such as MIG or time-slicing implement the share. They are not the policy. Production and training usually want whole cards.

Cap concurrent memory, not only slice count. Explore OneSource Cloud’s AI orchestration platform when multiple teams must request fractions against a visible quota on exclusive U.S. GPUs.