{ "schema_version": 1, "recorded_at": "2026-10-31T11:24:30+01:00", "purpose": "Débit de génération sur un prompt synthétique identique, trois répétitions par modèle. Ne mesure ni la qualité ni la durée d’une tâche complète d’agent.", "ollama": { "version": "0.40.13" }, "gpu_before": { "at": "2026-10-31T11:24:30+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 2.0, "utilization.gpu": 0.0, "temperature.gpu": 48.0, "power.draw": 28.49, "power.limit": 350.0 }, "request": { "prompt": "Rédige en français une analyse technique détaillée d'au moins 800 mots sur la sauvegarde d'un homelab auto-hébergé : services Docker, base PostgreSQL, fichiers, sauvegardes hors site, chiffrement, restauration et vérifications. Utilise des paragraphes explicatifs et des exemples concrets. Commence directement par le contenu, sans annoncer de plan.", "think": false, "stream": false, "keep_alive": "5m", "options": { "num_ctx": 65536, "num_predict": 256, "num_batch": 512, "temperature": 0, "seed": 42, "draft_num_predict": 0 } }, "protocol": "Avant chaque essai : décharger le modèle, le précharger sans prompt, puis mesurer une génération de 256 tokens. Le chargement et le traitement du prompt sont exclus du débit de génération.", "units": { "durations": "nanoseconds", "gpu_memory": "MiB", "throughput": "tokens/second" }, "models": [ { "tag": { "name": "qwen3.8:27b", "model": "qwen3.8:27b", "modified_at": "2026-08-19T00:33:37.333872424+02:00", "size": 17741872154, "digest": "22130167c4c20e20c7b71454612966ca8e8171e9b3cc8ab6ce8aa6cbfec79643", "details": { "parent_model": "qwen3.8:27b-q4_K_M", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "27.3B", "quantization_level": "Q4_K_M", "context_length": 262144, "embedding_length": 5120, "runner": "ggml" }, "capabilities": [ "completion", "vision", "tools", "thinking" ] }, "trials": [ { "repeat": 1, "at": "2026-10-31T11:24:39+01:00", "wall_seconds": 6.1031064910348505, "gpu_before": { "at": "2026-10-31T11:24:33+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 20999.0, "utilization.gpu": 45.0, "temperature.gpu": 63.0, "power.draw": 174.27, "power.limit": 350.0 }, "gpu_after": { "at": "2026-10-31T11:24:39+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 21009.0, "utilization.gpu": 97.0, "temperature.gpu": 74.0, "power.draw": 347.73, "power.limit": 350.0 }, "running_model": { "name": "qwen3.8:27b", "model": "qwen3.8:27b", "size": 20749284801, "digest": "22130167c4c20e20c7b71454612966ca8e8171e9b3cc8ab6ce8aa6cbfec79643", "details": { "parent_model": "", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "27.3B", "quantization_level": "Q4_K_M" }, "expires_at": "2026-10-31T11:29:39.707235748+01:00", "size_vram": 20749284801, "context_length": 65536, "runner": "llamacpp" }, "generation": { "model": "qwen3.8:27b", "done": true, "done_reason": "length", "load_duration": 795557, "total_duration": 6101183016, "prompt_eval_count": 96, "prompt_eval_duration": 141029000, "prompt_eval_cached_count": 0, "eval_count": 256, "eval_duration": 5957683000 }, "response_sha256": "9f740e4aed2471180b141178a5a8e06b706da74aa5603c8f14ea0d8a8c4bdd69", "generation_tokens_per_second": 42.96972497529661 }, { "repeat": 2, "at": "2026-10-31T11:24:49+01:00", "wall_seconds": 6.109854422044009, "gpu_before": { "at": "2026-10-31T11:24:43+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 20999.0, "utilization.gpu": 44.0, "temperature.gpu": 65.0, "power.draw": 178.57, "power.limit": 350.0 }, "gpu_after": { "at": "2026-10-31T11:24:49+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 21009.0, "utilization.gpu": 98.0, "temperature.gpu": 74.0, "power.draw": 340.58, "power.limit": 350.0 }, "running_model": { "name": "qwen3.8:27b", "model": "qwen3.8:27b", "size": 20749284801, "digest": "22130167c4c20e20c7b71454612966ca8e8171e9b3cc8ab6ce8aa6cbfec79643", "details": { "parent_model": "", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "27.3B", "quantization_level": "Q4_K_M" }, "expires_at": "2026-10-31T11:29:49.407651171+01:00", "size_vram": 20749284801, "context_length": 65536, "runner": "llamacpp" }, "generation": { "model": "qwen3.8:27b", "done": true, "done_reason": "length", "load_duration": 645581, "total_duration": 6107971735, "prompt_eval_count": 96, "prompt_eval_duration": 138829000, "prompt_eval_cached_count": 0, "eval_count": 256, "eval_duration": 5966831000 }, "response_sha256": "9f740e4aed2471180b141178a5a8e06b706da74aa5603c8f14ea0d8a8c4bdd69", "generation_tokens_per_second": 42.903846279541014 }, { "repeat": 3, "at": "2026-10-31T11:24:59+01:00", "wall_seconds": 6.1136000249534845, "gpu_before": { "at": "2026-10-31T11:24:53+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 20999.0, "utilization.gpu": 35.0, "temperature.gpu": 62.0, "power.draw": 177.75, "power.limit": 350.0 }, "gpu_after": { "at": "2026-10-31T11:24:59+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 21009.0, "utilization.gpu": 98.0, "temperature.gpu": 73.0, "power.draw": 340.61, "power.limit": 350.0 }, "running_model": { "name": "qwen3.8:27b", "model": "qwen3.8:27b", "size": 20749284801, "digest": "22130167c4c20e20c7b71454612966ca8e8171e9b3cc8ab6ce8aa6cbfec79643", "details": { "parent_model": "", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "27.3B", "quantization_level": "Q4_K_M" }, "expires_at": "2026-10-31T11:29:59.472109533+01:00", "size_vram": 20749284801, "context_length": 65536, "runner": "llamacpp" }, "generation": { "model": "qwen3.8:27b", "done": true, "done_reason": "length", "load_duration": 2615535, "total_duration": 6110995476, "prompt_eval_count": 96, "prompt_eval_duration": 141147000, "prompt_eval_cached_count": 0, "eval_count": 256, "eval_duration": 5964653000 }, "response_sha256": "9f740e4aed2471180b141178a5a8e06b706da74aa5603c8f14ea0d8a8c4bdd69", "generation_tokens_per_second": 42.91951266905217 } ], "median_generation_tokens_per_second": 42.91951266905217 }, { "tag": { "name": "bonsai-2:27b", "model": "bonsai-2:27b", "modified_at": "2026-10-31T10:24:55.232889638+01:00", "size": 8255266275, "digest": "f3c4f42086593225d8be9bb4270658d0e05f5232fb2b5e8faa49ef74d2cc843a", "details": { "parent_model": "bonsai-2:27b-q2_0", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "26.9B", "quantization_level": "Q2_0", "context_length": 262144, "embedding_length": 5120, "runner": "llamacpp" }, "capabilities": [ "completion", "vision", "tools", "thinking" ] }, "trials": [ { "repeat": 1, "at": "2026-10-31T11:25:06+01:00", "wall_seconds": 3.984593156957999, "gpu_before": { "at": "2026-10-31T11:25:02+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 12559.0, "utilization.gpu": 22.0, "temperature.gpu": 62.0, "power.draw": 174.44, "power.limit": 350.0 }, "gpu_after": { "at": "2026-10-31T11:25:06+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 12585.0, "utilization.gpu": 96.0, "temperature.gpu": 72.0, "power.draw": 325.08, "power.limit": 350.0 }, "running_model": { "name": "bonsai-2:27b", "model": "bonsai-2:27b", "size": 12195830168, "digest": "f3c4f42086593225d8be9bb4270658d0e05f5232fb2b5e8faa49ef74d2cc843a", "details": { "parent_model": "", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "26.9B", "quantization_level": "Q2_0" }, "expires_at": "2026-10-31T11:30:06.459774745+01:00", "size_vram": 12195830168, "context_length": 65536, "runner": "llamacpp" }, "generation": { "model": "bonsai-2:27b", "done": true, "done_reason": "length", "load_duration": 419095, "total_duration": 3982552211, "prompt_eval_count": 96, "prompt_eval_duration": 163426000, "prompt_eval_cached_count": 0, "eval_count": 256, "eval_duration": 3817106000 }, "response_sha256": "2ae435df17e63ef756e642d435657fff4198638b76f46fc5edc24825c617d2ae", "generation_tokens_per_second": 67.06651583686698 }, { "repeat": 2, "at": "2026-10-31T11:25:13+01:00", "wall_seconds": 3.985477131907828, "gpu_before": { "at": "2026-10-31T11:25:09+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 12559.0, "utilization.gpu": 14.0, "temperature.gpu": 61.0, "power.draw": 173.32, "power.limit": 350.0 }, "gpu_after": { "at": "2026-10-31T11:25:13+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 12585.0, "utilization.gpu": 96.0, "temperature.gpu": 73.0, "power.draw": 327.73, "power.limit": 350.0 }, "running_model": { "name": "bonsai-2:27b", "model": "bonsai-2:27b", "size": 12195830168, "digest": "f3c4f42086593225d8be9bb4270658d0e05f5232fb2b5e8faa49ef74d2cc843a", "details": { "parent_model": "", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "26.9B", "quantization_level": "Q2_0" }, "expires_at": "2026-10-31T11:30:13.408744621+01:00", "size_vram": 12195830168, "context_length": 65536, "runner": "llamacpp" }, "generation": { "model": "bonsai-2:27b", "done": true, "done_reason": "length", "load_duration": 402472, "total_duration": 3983626533, "prompt_eval_count": 96, "prompt_eval_duration": 165226000, "prompt_eval_cached_count": 0, "eval_count": 256, "eval_duration": 3816350000 }, "response_sha256": "2ae435df17e63ef756e642d435657fff4198638b76f46fc5edc24825c617d2ae", "generation_tokens_per_second": 67.0798013809006 }, { "repeat": 3, "at": "2026-10-31T11:25:20+01:00", "wall_seconds": 3.9983374520670623, "gpu_before": { "at": "2026-10-31T11:25:16+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 12559.0, "utilization.gpu": 44.0, "temperature.gpu": 61.0, "power.draw": 173.43, "power.limit": 350.0 }, "gpu_after": { "at": "2026-10-31T11:25:20+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 12585.0, "utilization.gpu": 96.0, "temperature.gpu": 72.0, "power.draw": 330.58, "power.limit": 350.0 }, "running_model": { "name": "bonsai-2:27b", "model": "bonsai-2:27b", "size": 12195830168, "digest": "f3c4f42086593225d8be9bb4270658d0e05f5232fb2b5e8faa49ef74d2cc843a", "details": { "parent_model": "", "format": "gguf", "family": "qwen35", "families": [ "qwen35" ], "parameter_size": "26.9B", "quantization_level": "Q2_0" }, "expires_at": "2026-10-31T11:30:20.367625935+01:00", "size_vram": 12195830168, "context_length": 65536, "runner": "llamacpp" }, "generation": { "model": "bonsai-2:27b", "done": true, "done_reason": "length", "load_duration": 381630, "total_duration": 3996412363, "prompt_eval_count": 96, "prompt_eval_duration": 171936000, "prompt_eval_cached_count": 0, "eval_count": 256, "eval_duration": 3822282000 }, "response_sha256": "2ae435df17e63ef756e642d435657fff4198638b76f46fc5edc24825c617d2ae", "generation_tokens_per_second": 66.97569671730135 } ], "median_generation_tokens_per_second": 67.06651583686698 } ], "comparison": { "generation_speed_ratio": 1.56261130814811, "generation_speed_increase_percent": 56.26113081481099 }, "gpu_after": { "at": "2026-10-31T11:25:20+01:00", "name": "NVIDIA GeForce RTX 3090", "driver_version": "550.163.01", "memory.total": 24576.0, "memory.used": 2.0, "utilization.gpu": 87.0, "temperature.gpu": 68.0, "power.draw": 309.51, "power.limit": 350.0 } }