{
  "title": "Edge Inference Benchmarks",
  "subtitle": "Compare throughput and latency across devices and model variants.",
  "model_column": "model",
  "model_family_column": "model_family",
  "model_link_prefix": "https://huggingface.co/",
  "optimized_org": "embedl",
  "filters": [
    {
      "column": "type",
      "label": "Modality"
    },
    {
      "column": "batch",
      "label": "Batch Size",
      "type": "number"
    },
    {
      "column": "device",
      "label": "Device",
      "value_labels": {
        "orin_nano": "Jetson Orin Nano Super",
        "orin_nano_super": "Jetson Orin Nano Super",
        "agx_orin": "Jetson AGX Orin",
        "agx_thor": "Jetson AGX Thor"
      }
    }
  ],
  "metrics": [
    {
      "column": "tps",
      "label": "Tokens / sec",
      "short": "TPS ↑",
      "higher_is_better": true,
      "description": "Tokens per second (higher is better). Number of output tokens generated per second during the decoding phase. "
    },
    {
      "column": "tpot",
      "label": "Time per Output Token (ms)",
      "short": "TPOT(ms) ↓",
      "higher_is_better": false,
      "description": "Time per output token in ms (lower is better). Average time (in milliseconds) required to generate one output token during decoding. Computed as TPOT = (last_token_ts - first_token_ts) / total_output_tokens."
    },
    {
      "column": "ttft",
      "label": "Time to First Token (ms)",
      "short": "TTFT(ms) ↓",
      "higher_is_better": false,
      "description": "Time to first token in ms (lower is better). Time from request submission to generation of the first output token. This includes vision encoding, prompt prefill, KV cache initialization."
    },
    {
      "column": "e2e",
      "label": "End-to-End Latency (sec)",
      "short": "E2E(s) ↓",
      "higher_is_better": false,
      "description": "End-to-end latency in seconds (lower is better). Total time from request submission to completion of the full generated response. This reflects real user-perceived latency."
    }
  ],
  "display_columns": [
    {
      "column": "res",
      "label": "RESOLUTION",
      "visible_when": {
        "type": [
          "video",
          "image"
        ]
      }
    },
    {
      "column": "fps",
      "label": "FPS",
      "type": "number",
      "visible_when": {
        "type": [
          "video"
        ]
      }
    },
    {
      "column": "frames",
      "label": "Frames",
      "type": "number",
      "visible_when": {
        "type": [
          "video"
        ]
      }
    }
  ],
  "chart": {
    "default_metric": "tps",
    "group_by": "device",
    "scenarios": [
      {
        "label": "Text",
        "match": {
          "type": "text"
        }
      },
      {
        "label": "Image · 1280×720",
        "match": {
          "type": "image",
          "res": "1280x720"
        }
      },
      {
        "label": "Video · 1280×720 · 4 FPS",
        "match": {
          "type": "video",
          "res": "1280x720",
          "fps": 4
        }
      }
    ]
  },
  "table_sort": [
    {
      "column": "res",
      "direction": "asc"
    },
    {
      "column": "fps",
      "direction": "desc"
    }
  ],
  "table_group_by": "model",
  "model_families": {
    "Cosmos-Reason2-2B": {
      "data_file": "data/Cosmos-Reason2.csv",
      "table_group_by": ["res", "fps"],
      "experiment_setup": {
        "agx_thor": "Measurement setup: NVIDIA vLLM 26.01, 256 tokens generated, 10 warm-up runs, averaged over 25 runs.",
        "agx_orin": "Measurement setup: NVIDIA AI IoT vLLM 0.14.0 tegra, 256 tokens generated, 10 warm-up runs, averaged over 25 runs.",
        "orin_nano": "Measurement setup: NVIDIA AI IoT vLLM 0.14.0 tegra, 256 tokens generated, 10 warm-up runs, averaged over 25 runs."
      }
    }
  }
}