Sizer Dataset And Administration
Sizer recommendations are based on a benchmark dataset compiled by Metrum and made available to Metrum Insights users.
This page is for Sizer users and customer admins who need to understand dataset coverage, catalog data, pricing assumptions, feature availability, and common administrative workflows.
What Sizer Uses
Sizer combines several kinds of product data:
| Data | Used for |
|---|---|
| Metrum benchmark dataset | Throughput, latency, and hardware behavior across supported model/runtime/hardware combinations. |
| Model and runtime metadata | Matching recommendations to model, framework, precision, and serving-method inputs. |
| GPU catalog data | Validating GPU models, VRAM, node sizing, and vendor filters. |
| Pricing data | Estimating capex, monthly opex, and amortized cost where available. |
| Owned inventory | Restricting a request to hardware you already have. |
When a field is left unspecified, Sizer keeps that dimension open. For example, a model-only request can search across available framework and precision coverage, while an inventory-only request can search for viable model/runtime combinations on the listed GPUs.
Dataset Coverage
The Metrum benchmark dataset is shared across Metrum Insights users. It is curated and expanded by Metrum as new models, runtimes, GPUs, and workload profiles become relevant.
Sizer can return useful recommendations from minimal input, as long as the request includes modality plus at least one sizing signal. Additional fields make the recommendation more specific to the question you are asking. Dataset coverage matters most when you pin the request to an exact model, runtime, GPU, workload shape, or combination of constraints.
If an exact combination is not covered yet, the result may include fewer exact-match candidates or a warning that explains the coverage gap. If that combination matters for your deployment decision, contact Metrum about adding benchmark coverage.
Feature Availability
The Sizer page appears when the feature is enabled for your Metrum Insights instance. Customer admins and API users can use these endpoints:
| Endpoint | Use it for |
|---|---|
GET /api/sizer/status | Workspace configuration: whether Sizer is enabled for the workspace. |
GET /api/sizer/health | Service readiness: whether the Sizer service is reachable right now. |
Most Sizer users do not need these endpoints. They are useful when confirming that Sizer is enabled or checking service readiness.
Pricing And Cost Assumptions
Cost fields depend on available pricing data. Recommendations can include:
- capex;
- monthly opex;
- amortized monthly cost;
- bill of materials details when available.
When pricing is estimated or incomplete, Sizer marks the source and includes warnings instead of treating the value as exact. Treat estimated cost output as guidance, not a purchase quote.
When A Request Cannot Be Answered
Common causes:
| Cause | What to try |
|---|---|
| Coverage, catalog, or pricing data is unavailable | Use supported GPU models from the Sizer inventory selector or GPU catalog, loosen model/runtime/hardware/workload filters, or contact Metrum about benchmark coverage, catalog support, or pricing data. |
| Request constraints are too narrow | Remove or relax a hardware, framework, precision, budget, latency, inventory, or throughput-target constraint and retry. |
Owned Inventory
Owned inventory is request-specific. Entering inventory in Sizer constrains that run of the tool; it does not change the shared hardware catalog or the Metrum benchmark dataset.
For inventory CSV import, use:
gpu_model,count
h200,16
b200,8
GPU models must match the catalog, and the number of user-owned GPUs must be a positive integer.