system_desc.json

The hardware and software description of the system under test. One per Pareto point.

Not written by the reference client

You supply it and drop it into each run folder before upload. See step 5.

The checker verifies that every point of a curve describes the same system.

Fields and template

The fields, and what each one means, are defined in rules §8.2. There are two ways to produce the file:

• Capture it with mlperf-sysinfo. It reads the machine under test and writes system_desc.json using its endpoints profile. How to install and run it is in its docs.
• Copy the template in §8.2.1, which includes the nesting of node_types and accelerator_info, and fill it in.

tps_utilization is computed

The checker recomputes it against your own curve (tps-utilization).

cooling sets your power overhead

The checker uses cooling to pick the overhead fraction for system_power.json: 0.30 for liquid-cooled, 0.50 for air-cooled. Leave it empty and power-descriptor fails.

No model_name here

The benchmark model name goes in point.yaml (§8.3). §8.2 no longer lists it, and the checker ignores a model_name left in this file.

disaggregated is truncated upstream

The field definition in the source data dictionary is cut off mid-sentence and carries an upstream TODO to verify the full text with the dictionary owner. Confirm the intended semantics before relying on it.

Provisioned power is not in this file. It goes in a separate, per-system system_power.json.

Last verified against: mlcommons/endpoints_policies@v1.0_rules_dev (d2d9da6) and mlcommons/endpoints-submission-cli@main (a42a056, v1.1.0.0), 2026-10-02.