# IRI Link Relation: `has-gpu`

**Relation URI:** `https://iri.science/rels/has-gpu`\
**CURIE:** `iri:has-gpu`\
**Status:** Draft\
**Version:** 1.0.0<br>
**Source representation type:** `urn:doe-iri:resource:compute:node`<br>
**Source resource type:** `urn:doe-iri:resource:compute:node`<br>
**Target representation type:** `urn:doe-iri:resource:compute:gpu`<br>
**Target resource type:** `urn:doe-iri:resource:compute:gpu`

This document defines the `iri:has-gpu` relationship used by the DOE-IRI compute resource model.

The canonical relation URI is `https://iri.science/rels/has-gpu`. With the
canonical IRI CURIE template `https://iri.science/rels/{rel}`, `iri:has-gpu`
expands to that URI. The relation URI identifies the link-relation semantics
and is distinct from any target representation profile.

## 1. Relationship Metadata

| Field | Definition |
|---|---|
| Relationship | `iri:has-gpu` |
| Relation URI | `https://iri.science/rels/has-gpu` |
| Semantic meaning | Indicates that the compute node contains, provides, or is associated with the identified GPU resource. |
| Source representation type | `urn:doe-iri:resource:compute:node` |
| Target representation type | `urn:doe-iri:resource:compute:gpu` |
| Cardinality | `0..*` targets from a compute-node resource. |
| Target stability | Static resource representation. The target identifies a GPU resource independent of current utilization, memory consumption, health, or allocation. |
| Authorization affects visibility | Yes. GPU-level topology MAY be omitted when the requester is not authorized to discover accelerator-level details. |
| Target classification | Resource |
| Relationship volatility | Relatively static hardware/topology relationship. Ordinary GPU operational changes do not change the relationship. |

## 2. Semantic Meaning

The `iri:has-gpu` relationship identifies GPU or accelerator resources associated with the source compute node.

GPU-specific characteristics such as device memory, programming interfaces, vendor, model, and vendor-defined architecture belong to the target GPU resource rather than being embedded into the relationship.

## 3. Source and Target Representation

The relationship MUST originate from:

```text
urn:doe-iri:resource:compute:node
```

and MUST target:

```text
urn:doe-iri:resource:compute:gpu
```

The target is a Resource representation, not an operation entry point or relationship resource.

## 4. Cardinality

A compute node MAY expose zero, one, or multiple GPU resources:

```text
Compute Node  -- iri:has-gpu -->  GPU
     1                0..*
```

A node with no GPU targets may be CPU-only, or the facility may have chosen not to expose individual accelerator topology. Clients SHOULD use the node's `gpu_count` attribute, when present, to distinguish these cases where necessary.

## 5. Static and Dynamic Semantics

The relationship describes relatively stable accelerator topology. It SHOULD NOT be added or removed solely because a GPU is allocated, idle, unhealthy, or temporarily unavailable.

Current utilization, allocated memory, temperature, health, workload assignment, and availability are outside the semantics of this relation. When represented, they are governed by the applicable IRI API contract and Resource Definition Profile.

## 6. Authorization and Visibility

Authorization MAY affect visibility of GPU-level topology. A provider MAY omit GPU targets while still exposing aggregate accelerator counts on the node or compute-system resource.

Absence of visible GPU targets MUST NOT be interpreted as evidence that no accelerators exist unless the applicable aggregate attributes also indicate zero.

## 7. HAL Representation

```json
{
  "_links": {
    "iri:has-gpu": [
      {
        "href": "/api/v2/status/resources/node-001-gpu-0",
        "profile": "https://iri.science/profiles/resource-definition/compute/gpu"
      },
      {
        "href": "/api/v2/status/resources/node-001-gpu-1",
        "profile": "https://iri.science/profiles/resource-definition/compute/gpu"
      }
    ]
  }
}
```

---

*DOE Integrated Research Infrastructure — Link Relation: has-gpu*