Context{}

Before selecting or designing a GPUaaS solution, we need a clear and explicit definition of what “GPUaaS” actually means for our organization.

Today, different technologies offer overlapping but incomplete capabilities around GPU scheduling, quotas, priorities, fairness, and preemption.

Without a shared requirements baseline, technology evaluation becomes subjective and architecture decisions are driven by assumptions rather than facts.

 

Objective{}

Create a structured GPUaaS requirements document that will serve as the single source of truth for evaluating and comparing candidate technologies.

 

The document must make it possible to clearly answer:

“What do we require?” and later “Which technology supports what, and how?”

 

Description{}

This story focuses on defining requirements only, not evaluation or implementation.

 

The output will be a requirements table that explicitly describes GPUaaS expectations in plain English and allows future marking of support status per technology.

 

Each requirement should be written in a way that enables a simple comparison matrix, for example:

Requirement → Supported / Partially Supported / Not Supported.

 

Scope{}

In scope:

 

• GPU scheduling semantics (priority, fairness, preemption)

• GPU ownership and quota models

• Namespace and team-level isolation

• Opportunistic usage behavior

• Observability, logging, and user-facing feedback

• GPU heterogeneity handling (types, models, constraints)

• Policy and configurability expectations

 

 

Out of scope:

 

• Choosing a specific GPUaaS technology

• Designing the final GPUaaS architecture

• Implementation or production rollout

 

 

Deliverables{}

A GPUaaS requirements document containing a table with (at minimum) the following columns:

 

• Category

• Requirement

• Keyword (MUST / SHOULD / MAY)

• What does it mean (plain English)

• Rationale / Use Case

 

 

The table must be written so that it can later be extended with:

 

• Technology name

• Support status

• Notes / gaps

 

 

DoD{}

This story is complete when:

 

• All GPUaaS requirements are documented in a structured table

• Each requirement is clearly explained in plain English

• Requirements are written in a way that enables direct comparison across technologies

• The document can be used as input for the technology evaluation epic that follows

Notes{}

This story intentionally defines what we need, not how to build it.

The output of this story is a prerequisite for any meaningful GPUaaS technology evaluation.