`system_desc.json`¶

The hardware and software description of the system under test. One per Pareto point, since policies PR #119 there is no shared per-system file.

Authored by you

Not written by the reference client. You author it and drop it into each run folder before upload. Authored in step 5.

The checker verifies that every point of a curve describes the same system (system-description-consistency).

Identity and classification¶

Field	Description
`division`	`Standardized`, `Serviced` or `RDI`
`system_name`	Your string describing the system under test
`shortened_system_name`	Shortened name, at most 20 characters
`system_availability_status`	`Available`, `Preview` or `RDI` at submission time
`model_name`	Benchmark model name — must match the round's supported model list
`max_supported_concurrency`	Your declared `C_max`
`endpoint_url`	URL or description of the endpoint under test

Scale and topology¶

Field	Description
`system_size`	Accelerators per node type, e.g. `"72 accelerators + 144 accelerators"` for a two-node-type system
`system_node_ensemble_count`	How many unique hardware+software combinations are in the system
`system_node_ensemble_total`	Total nodes — the sum of all `number_of_nodes`
`system_node_ensemble_id`	Identifies a unique node type within the system
`number_of_nodes`	How many nodes of this type

Host¶

Field	Description
`host_processor_model_name`	Host processor model
`host_processors_per_node`	Host processors per node
`host_processor_core_count`	CPU cores per processor — optional, but at least one of core/vCPU count must be present
`host_processor_vcpu_count`	vCPUs per processor — same condition
`host_memory_capacity`	Total memory for all host processors, not per-processor
`host_memory_configuration`	DIMM count, memory type (DDR5, LPDDR4) and speed
`host_network_card_count`	Number and type of networking cards, with speeds
`host_networking`	Protocol — InfiniBand, Ethernet
`host_storage_capacity`	Total storage for the node
`host_storage_type`	Storage type

Accelerator¶

Field	Description
`accelerator_model_name`	Accelerator model
`accelerators_per_node`	Accelerators per node
`accelerator_memory_capacity`	Memory per accelerator
`accelerator_memory_type`	Accelerator memory type
`accelerator_host_interconnect`	Link between accelerator and host processors
`accelerator_interconnect`	Link between accelerators, where `accelerators_per_node > 1`

Software¶

Field	Description
`serving_framework`	SGLang, vLLM, etc.
`inference_backend`	Vendor stack components
`driver`	Driver and version for any accelerators
`operating_system`	OS
`filesystem`	Filesystem
`container_link`	Link to the submission container
`other_software_stack`	Other performance-relevant software, free-form
`sw_notes`	Supplementary software notes, free-form

Parallelism and configuration¶

Field	Description
`tensor_parallel`	TP=N splits weight matrices N ways; each partition holds 1/N of each layer. Attention head count generally must divide by N. TP=1 means no partitioning
`expert_parallel`	EP=N splits experts into N groups on different accelerators, routing tokens to the right group. MoE models only. EP=1 means no partitioning
`pipeline_parallel`	PP=N splits layers into N sequential stages; an inference passes through all stages. PP=1 means no partitioning
`data_parallel`	DP=N replicates the model N times and distributes requests. DP=1 means no replication
`disaggregated`	Whether the system is disaggregated (`> 1`)
`batch`	Maximum batch size
`node_config`	Configuration of nodes/processors, in enough detail to reproduce the submission
`config_summary`	Concatenation of `disaggregated`, `tensor_parallel`, `pipeline_parallel`, `expert_parallel`, `data_parallel` where each is > 1, plus `config_summary_notes`
`config_summary_notes`	Free-form, for anything the concatenated fields miss
`link_config`	Link to full configuration logs for the run

Other¶

Field	Description
`cooling`	Liquid, air-only, or passive
`other_hardware`	Other performance-relevant hardware, free-form
`hw_notes`	Supplementary hardware notes, free-form
`tps_utilization`	`reported_system_tps / max(reported_system_tps across all runs)`

tps_utilization is computed, not chosen

The checker recomputes it against your own curve (tps-utilization). You cannot fill it in until every point has run.

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.

Template¶

{
  "division": "",
  "system_name": "",
  "shortened_system_name": "",
  "system_availability_status": "",
  "system_size": "",
  "system_node_ensemble_count": 0,
  "system_node_ensemble_total": 0,
  "endpoint_url": "",
  "serving_framework": "",
  "node_types": [
    {
      "system_node_ensemble_id": 0,
      "number_of_nodes": 0,
      "host_processor_model_name": "",
      "host_processors_per_node": 0,
      "host_processor_core_count": 0,
      "host_processor_vcpu_count": 0,
      "host_memory_capacity": "",
      "host_memory_configuration": "",
      "accelerator_info": [
        {
          "accelerator_model_name": "",
          "accelerators_per_node": 0,
          "accelerator_memory_capacity": "",
          "accelerator_memory_type": "",
          "accelerator_interconnect": "",
          "accelerator_host_interconnect": ""
        }
      ],
      "host_network_card_count": "",
      "host_networking": "",
      "host_storage_capacity": "",
      "host_storage_type": "",
      "other_hardware": "",
      "cooling": "",
      "hw_notes": "",
      "inference_backend": "",
      "driver": "",
      "operating_system": "",
      "filesystem": "",
      "container_link": "",
      "other_software_stack": "",
      "sw_notes": ""
    }
  ],
  "node_config": "",
  "disaggregated": 0,
  "expert_parallel": 0,
  "tensor_parallel": 0,
  "pipeline_parallel": 0,
  "data_parallel": 0,
  "batch": 0,
  "config_summary": "",
  "config_summary_notes": "",
  "link_config": "",
  "tps_utilization": 0
}

Checker rules that read this file¶

Rule	Checks
`system-description-present`	Every point has a `system_desc.json`
`system-description-valid`	It parses against the `SystemDescription` schema
`system-description-consistency`	Every point of a curve describes the same system
`model-name-valid`	`model_name` is one of the round's supported models
`model-name-consistency`	It matches the results directory name
`max-concurrency-declared`	`max_supported_concurrency` present and > 32
`tps-utilization`	Equals `system_tps / max(system_tps)` over the point's own curve

Field name drift

The rules refer to the file as system_desc_id.json in places and to a benchmark_model field in the result-ID definition, while the tooling uses system_desc.json and model_name. The tooling spelling is what the checker reads.

Last verified against: mlcommons/endpoints_policies@v1.0_rules_dev (a7ec3cc) and mlcommons/endpoints-submission-cli@main (f48ca84), 2026-09-19.

system_desc.json¶