Initial upload from tokenwatt.io published data

README.md

@@ -0,0 +1,12 @@
+---
+license: cc-by-4.0
+pretty_name: "TokenWatt inference reports"
+tags: [llm-inference, vllm, benchmark, h100, goodput, tokens-per-joule]
+---
+
+# TokenWatt inference reports
+
+Published TokenWatt Bench reports, as shown on [tokenwatt.io/reports](https://www.tokenwatt.io/reports.html). Each report gives SLO goodput (the highest request rate meeting p99 TTFT and TPOT targets), tokens per joule, memory-bandwidth utilization and an accuracy gate for each configuration measured. Method: [methodology](https://www.tokenwatt.io/methodology.html).
+
+- `h100-qwen3-30b-a3b-fp8-vllm-0-31.json` — H100 SXM, Qwen3-30B-A3B-FP8, vLLM 0.31.0, full mode: 3,996 output tok/s at 9.5 req/s within SLO
+- `lab-h100-qwen3-30b-a3b-kv-cache.json` — BF16 vs FP8 KV cache on the same setup ([Lab notes #1](https://www.tokenwatt.io/blog/lab-notes-1-h100-fp8-kv-cache.html))

h100-qwen3-30b-a3b-fp8-vllm-0-31.json

@@ -0,0 +1,164 @@
+{
+  "id": "h100-qwen3-30b-a3b-fp8-vllm-0-31",
+  "status": "published",
+  "title": "Qwen3-30B-A3B FP8 on 1\u00d7 H100 SXM with vLLM 0.31.0",
+  "published": "2026-10-06",
+  "summary": "Full-mode measurement on a Lambda Cloud 1\u00d7 H100 SXM5 instance: 10-minute windows, 5 repeats at the SLO boundary, GSM8K 5-shot on 200 items \u00d7 3 seeds. Default vLLM configuration with a BF16 KV cache. Energy is GPU board power.",
+  "hardware": {
+    "vendor": "NVIDIA",
+    "accelerator": "NVIDIA H100 80GB HBM3",
+    "count": 1,
+    "rated_bw_tbs_per_gpu": 3.35,
+    "driver": "580.105.08",
+    "host": "Lambda Cloud instance"
+  },
+  "software": {
+    "engine": "vLLM",
+    "engine_version": "0.31.0",
+    "kernels": "vllm/vllm-openai:latest"
+  },
+  "model": {
+    "name": "Qwen3-30B-A3B-FP8",
+    "shape": "MoE",
+    "quantization": "FP8 weights, BF16 KV cache",
+    "parallelism": "1 replica \u00d7 TP1"
+  },
+  "slo": {
+    "ttft_p99_ms": 2000.0,
+    "tpot_p99_ms": 50.0
+  },
+  "workload": {
+    "trace": "chat-mix-v1",
+    "input_tokens": "median 1,024 \u00b7 p90 4,096",
+    "output_tokens": "median 256 \u00b7 p90 1,024",
+    "arrival": "Poisson, rate swept to SLO boundary"
+  },
+  "power": {
+    "source": "GPU board power sum (NVIDIA)",
+    "kind": "board",
+    "sampling_hz": 1,
+    "window": "600 s steady state"
+  },
+  "accuracy": {
+    "suite": "gsm8k-5shot",
+    "threshold_pts": 1.0
+  },
+  "runs": 5,
+  "variance_band_pct": 2.2,
+  "fleet": {
+    "cards": 64
+  },
+  "configs": [
+    {
+      "key": "baseline",
+      "label": "vllm-default",
+      "engine": "vLLM 0.31.0 \u00b7 KV BF16",
+      "mbu_pct": 58.1,
+      "node_power_kw": 0.601,
+      "goodput_tok_s": 3995.9,
+      "ttft_p99_ms": 327.0,
+      "tpot_p99_ms": 37.3,
+      "slo_pass_count": 5,
+      "repeats": 5,
+      "throttle_counts": {
+        "sw_power_cap": 1870
+      },
+      "accuracy": {
+        "score": 94.17,
+        "delta_pts": 0.0,
+        "spread_pts": 0.5
+      }
+    }
+  ],
+  "not_covered": [
+    "Whole-node power (no BMC access on this host)",
+    "Context lengths beyond the chat-mix-v1 trace"
+  ],
+  "reproduce": "tokenwatt run --spec lambda-qwen3-30b-a3b-fp8-tp1-dp1.spec.yaml --label 'vllm-default' --role baseline --out runs/vllm-default.json\n\ntokenwatt report runs/vllm-default.json --title 'Qwen3-30B-A3B FP8 on 1\u00d7 H100 SXM with vLLM 0.31.0'",
+  "spec": {
+    "name": "lambda-qwen3-30b-a3b-fp8-tp1-dp1",
+    "hardware": {
+      "adapter": "nvidia",
+      "rated_bw_tbs_per_gpu": null,
+      "gpus": null,
+      "host": "Lambda Cloud instance"
+    },
+    "engine": {
+      "endpoint": "http://127.0.0.1:8000/v1",
+      "name": "vLLM",
+      "version": "0.31.0",
+      "kernels": "vllm/vllm-openai:latest",
+      "api": "completions",
+      "model": "Qwen/Qwen3-30B-A3B-FP8",
+      "ignore_eos": true,
+      "api_key_env": null
+    },
+    "model": {
+      "name": "Qwen3-30B-A3B-FP8",
+      "config": "/home/ubuntu/tokenwatt-bench/models/Qwen__Qwen3-30B-A3B-FP8/config.json",
+      "weight_dtype": "fp8",
+      "kv_dtype": "bf16",
+      "quantization": "",
+      "replicas": 1,
+      "tensor_parallel": 1,
+      "vocab_size": 151936,
+      "geometry": {}
+    },
+    "slo": {
+      "ttft_p99_ms": 2000.0,
+      "tpot_p99_ms": 50.0
+    },
+    "workload": {
+      "trace": "chat-mix-v1",
+      "input_tokens": null,
+      "output_tokens": null,
+      "seed": 0
+    },
+    "measurement": {
+      "warmup_s": 120,
+      "window_s": 600,
+      "search_window_s": 120.0,
+      "repeats": 5,
+      "request_timeout_s": 600,
+      "max_concurrency": 8192
+    },
+    "sweep": {
+      "rates": null,
+      "start_rps": 9.0,
+      "growth": 1.25,
+      "max_rps": 2000,
+      "bisect_steps": 4,
+      "backoff": 0.9,
+      "max_backoffs": 4
+    },
+    "accuracy": {
+      "suite": "/home/ubuntu/tokenwatt-bench/suites/gsm8k-200.jsonl",
+      "name": "gsm8k-5shot",
+      "scorer": "numeric",
+      "max_drop_pts": 1.0,
+      "seeds": 3,
+      "max_tokens": 384,
+      "stop": [
+        "\n\nQuestion:"
+      ],
+      "concurrency": 32,
+      "limit": 200
+    },
+    "power": {
+      "source": "board",
+      "sampling_hz": 1,
+      "url": null,
+      "username": null,
+      "password_env": null,
+      "verify_tls": false,
+      "chassis": null,
+      "host": null,
+      "command": null,
+      "field": null
+    },
+    "not_covered": [
+      "Whole-node power (no BMC access on this host)",
+      "Context lengths beyond the chat-mix-v1 trace"
+    ]
+  }
+}

lab-h100-qwen3-30b-a3b-kv-cache.json

@@ -0,0 +1,187 @@
+{
+  "id": "lab-h100-qwen3-30b-a3b-kv-cache",
+  "status": "draft",
+  "title": "Qwen3-30B-A3B FP8 on 1\u00d7 H100: BF16 vs FP8 KV cache",
+  "published": "2026-10-06",
+  "summary": "Validation run on a Lambda Cloud 1\u00d7 H100 SXM5 instance with vLLM 0.31.0. Quick mode: 120 s windows, 2 repeats, 50 GSM8K items with one seed. FP8 KV cache raised SLO goodput and tokens per joule but failed the accuracy gate on this small suite.",
+  "hardware": {
+    "vendor": "NVIDIA",
+    "accelerator": "NVIDIA H100 80GB HBM3",
+    "count": 1,
+    "rated_bw_tbs_per_gpu": 3.35,
+    "driver": "580.105.08",
+    "host": "Lambda Cloud instance"
+  },
+  "software": {
+    "engine": "vLLM",
+    "engine_version": "0.31.0",
+    "kernels": "vllm/vllm-openai:latest"
+  },
+  "model": {
+    "name": "Qwen3-30B-A3B-FP8",
+    "shape": "MoE",
+    "quantization": "FP8 weights; KV cache varies by configuration",
+    "parallelism": "1 replica \u00d7 TP1"
+  },
+  "slo": {
+    "ttft_p99_ms": 2000.0,
+    "tpot_p99_ms": 50.0
+  },
+  "workload": {
+    "trace": "chat-mix-v1",
+    "input_tokens": "median 1,024 \u00b7 p90 4,096",
+    "output_tokens": "median 256 \u00b7 p90 1,024",
+    "arrival": "Poisson, rate swept to SLO boundary"
+  },
+  "power": {
+    "source": "GPU board power sum (NVIDIA)",
+    "kind": "board",
+    "sampling_hz": 1,
+    "window": "120 s steady state"
+  },
+  "accuracy": {
+    "suite": "gsm8k-5shot",
+    "threshold_pts": 1.0
+  },
+  "runs": 2,
+  "variance_band_pct": 1.0,
+  "fleet": {
+    "cards": 64
+  },
+  "configs": [
+    {
+      "key": "baseline",
+      "label": "vllm-default",
+      "engine": "vLLM 0.31.0 \u00b7 KV BF16",
+      "mbu_pct": 56.6,
+      "node_power_kw": 0.637,
+      "goodput_tok_s": 4572.3,
+      "ttft_p99_ms": 340.0,
+      "tpot_p99_ms": 44.6,
+      "slo_pass_count": 2,
+      "repeats": 2,
+      "throttle_counts": {
+        "sw_power_cap": 199
+      },
+      "accuracy": {
+        "score": 96.0,
+        "delta_pts": 0.0,
+        "spread_pts": 0.0
+      }
+    },
+    {
+      "key": "tuned",
+      "label": "vllm-fp8kv",
+      "engine": "vLLM 0.31.0 \u00b7 KV FP8",
+      "mbu_pct": 45.3,
+      "node_power_kw": 0.616,
+      "goodput_tok_s": 5056.9,
+      "ttft_p99_ms": 352.0,
+      "tpot_p99_ms": 43.7,
+      "slo_pass_count": 2,
+      "repeats": 2,
+      "throttle_counts": {
+        "sw_power_cap": 133
+      },
+      "accuracy": {
+        "score": 88.0,
+        "delta_pts": -8.0,
+        "spread_pts": 0.0
+      }
+    }
+  ],
+  "not_covered": [
+    "Whole-node power (cloud VM: no BMC access)",
+    "Context lengths beyond the chat-mix-v1 trace",
+    "Quick mode: short windows, few repeats \u2014 validation only, not for publishing",
+    "Whole-node power: energy figures are GPU board power only (no CPUs, memory, NICs, fans or PSU losses)"
+  ],
+  "reproduce": "tokenwatt run --spec lambda-qwen3-30b-a3b-fp8-tp1-dp1.spec.yaml --label 'vllm-default' --role baseline --out runs/vllm-default.json\ntokenwatt run --spec lambda-qwen3-30b-a3b-fp8-tp1-dp1.spec.yaml --label 'vllm-fp8kv' --role tuned --out runs/vllm-fp8kv.json\n\ntokenwatt report runs/vllm-default.json runs/vllm-fp8kv.json --title 'Qwen3-30B-A3B FP8 on 1\u00d7 H100: BF16 vs FP8 KV cache'",
+  "spec": {
+    "name": "lambda-qwen3-30b-a3b-fp8-tp1-dp1",
+    "hardware": {
+      "adapter": "nvidia",
+      "rated_bw_tbs_per_gpu": null,
+      "gpus": null,
+      "host": "Lambda Cloud instance"
+    },
+    "engine": {
+      "endpoint": "http://127.0.0.1:8000/v1",
+      "name": "vLLM",
+      "version": "0.31.0",
+      "kernels": "vllm/vllm-openai:latest",
+      "api": "completions",
+      "model": "Qwen/Qwen3-30B-A3B-FP8",
+      "ignore_eos": true,
+      "api_key_env": null
+    },
+    "model": {
+      "name": "Qwen3-30B-A3B-FP8",
+      "config": "/home/ubuntu/tokenwatt-bench/models/Qwen__Qwen3-30B-A3B-FP8/config.json",
+      "weight_dtype": "fp8",
+      "kv_dtype": "bf16",
+      "quantization": "",
+      "replicas": 1,
+      "tensor_parallel": 1,
+      "vocab_size": 151936,
+      "geometry": {}
+    },
+    "slo": {
+      "ttft_p99_ms": 2000.0,
+      "tpot_p99_ms": 50.0
+    },
+    "workload": {
+      "trace": "chat-mix-v1",
+      "input_tokens": null,
+      "output_tokens": null,
+      "seed": 0
+    },
+    "measurement": {
+      "warmup_s": 45,
+      "window_s": 120,
+      "search_window_s": 60,
+      "repeats": 2,
+      "request_timeout_s": 300,
+      "max_concurrency": 8192
+    },
+    "sweep": {
+      "rates": null,
+      "start_rps": 2,
+      "growth": 2,
+      "max_rps": 2000,
+      "bisect_steps": 3,
+      "backoff": 0.9,
+      "max_backoffs": 4
+    },
+    "accuracy": {
+      "suite": "/home/ubuntu/tokenwatt-bench/suites/gsm8k-200.jsonl",
+      "name": "gsm8k-5shot",
+      "scorer": "numeric",
+      "max_drop_pts": 1.0,
+      "seeds": 1,
+      "max_tokens": 384,
+      "stop": [
+        "\n\nQuestion:"
+      ],
+      "concurrency": 32,
+      "limit": 50
+    },
+    "power": {
+      "source": "board",
+      "sampling_hz": 1,
+      "url": null,
+      "username": null,
+      "password_env": null,
+      "verify_tls": false,
+      "chassis": null,
+      "host": null,
+      "command": null,
+      "field": null
+    },
+    "not_covered": [
+      "Whole-node power (cloud VM: no BMC access)",
+      "Context lengths beyond the chat-mix-v1 trace",
+      "Quick mode: short windows, few repeats \u2014 validation only, not for publishing"
+    ]
+  }
+}