{
  "run_date": "2026-08-29",
  "title": "Qwen/Qwen3.8-27B on AMD Instinct MI300X: aggregated single-replica controls",
  "experiment_class": "single-replica-control",
  "replicas_per_arm": 1,
  "claim_scope": "Per-replica capacity and wrapper overhead only; not multi-replica routing, scaling, high availability, or failover.",
  "model": "Qwen/Qwen3.8-27B",
  "hardware": "1x AMD Instinct MI300X (192 GiB) per framework",
  "benchmark": {
    "tool": "NVIDIA AIPerf",
    "version": "0.12.0",
    "endpoint_type": "chat",
    "streaming": true,
    "concurrency": 4,
    "request_count": 4,
    "input_length_stddev": 0,
    "output_length_stddev": 0,
    "random_seed": 42,
    "forced_output_length": true
  },
  "systems": {
    "dynamo": {
      "label": "Dynamo aggregated vLLM",
      "dynamo_version": "1.4.2",
      "vllm_version": "0.28.1rc1.dev43+g6f7df92a8",
      "request_path": "Dynamo frontend -> Dynamo vLLM worker"
    },
    "llmd": {
      "label": "llm-d aggregated vLLM",
      "vllm_version": "0.28.0",
      "request_path": "Envoy -> llm-d EPP -> vLLM"
    }
  },
  "results": [
    {
      "scenario": "1K/8K",
      "isl_tokens": 1024,
      "osl_tokens": 8192,
      "dynamo": {"output_tokens_per_second": 132.0631, "ttft_ms": 546.6269, "itl_ms": 30.2190, "latency_ms": 248063.0795},
      "llmd": {"output_tokens_per_second": 83.1738, "ttft_ms": 804.0643, "itl_ms": 47.9965, "latency_ms": 393907.1268}
    },
    {
      "scenario": "8K/8K",
      "isl_tokens": 8192,
      "osl_tokens": 8192,
      "dynamo": {"output_tokens_per_second": 83.7114, "ttft_ms": 5238.6819, "itl_ms": 47.1426, "latency_ms": 391372.1119},
      "llmd": {"output_tokens_per_second": 83.2697, "ttft_ms": 4882.5749, "itl_ms": 47.4375, "latency_ms": 393443.2072}
    },
    {
      "scenario": "8K/1K",
      "isl_tokens": 8192,
      "osl_tokens": 1024,
      "dynamo": {"output_tokens_per_second": 95.9692, "ttft_ms": 390.8985, "itl_ms": 41.3060, "latency_ms": 42646.9834},
      "llmd": {"output_tokens_per_second": 105.2785, "ttft_ms": 411.1833, "itl_ms": 37.6142, "latency_ms": 38890.4590}
    }
  ],
  "interpretation": "At concurrency 4, Dynamo leads the decode-heavy 1K/8K scenario by 58.8%, the two systems are within 0.5% for 8K/8K throughput, and llm-d leads 8K/1K by 9.7%. Four requests per point make this a controlled lab snapshot, not a statistically robust capacity study."
}