[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f6t0astpzag7h":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":36,"bodyZh":52,"bodyEn":53,"sourceName":54,"sourceUrl":55,"updatedAt":56},"edge-ai-jetson-snapdragon-2026-10-16","GUIDE","2026-10-16T14:00:00.000Z","边缘 AI 实测：Jetson AGX Orin 110 tok\u002Fs @ 37W，每瓦 3 tok\u002Fs 创纪录","Edge AI measurements: Jetson AGX Orin 110 tok\u002Fs @ 37W, 3 tok\u002Fs per watt record","10-16 批量 108 条边缘 \u002F 移动设备实测：Jetson AGX Orin 64GB 平均 110.96 tok\u002Fs（37.5W），Snapdragon X Elite 80.82（22.5W），Snapdragon 7 Gen 4 78.59（21.2W）。**每瓦 tok\u002Fs：Jetson 2.96 创 DB 记录**。","10-16 batch 108 edge \u002F mobile device measurements: Jetson AGX Orin 64GB avg 110.96 tok\u002Fs (37.5W), Snapdragon X Elite 80.82 (22.5W), Snapdragon 7 Gen 4 78.59 (21.2W). **Per-watt tok\u002Fs: Jetson 2.96 — DB record**.",[12,15,18,21,24,27,30,33],{"id":13,"name":14},48,"Qwen3.5-4B",{"id":16,"name":17},49,"Qwen3.5-9B",{"id":19,"name":20},50,"Qwen3.5-27B",{"id":22,"name":23},51,"Qwen3.5-0.8B",{"id":25,"name":26},55,"Qwen3-1.7B",{"id":28,"name":29},60,"Qwen2.5-Omni-3B",{"id":31,"name":32},62,"Qwen3-Embedding-0.6B",{"id":34,"name":35},66,"gemma-4-E2B-it",[37,40,43,46,49],{"id":38,"name":39},13,"NVIDIA RTX 5090",{"id":41,"name":42},28,"Apple M5 Max",{"id":44,"name":45},76,"NVIDIA Jetson AGX Orin 64GB",{"id":47,"name":48},104,"Qualcomm Snapdragon X Elite",{"id":50,"name":51},105,"Qualcomm Snapdragon 7 Gen 4","# 边缘 AI 实测：Jetson AGX Orin 110 tok\u002Fs @ 37W，每瓦 3 tok\u002Fs 创纪录\n\n10-16 批量 108 条边缘 \u002F 移动设备 L2 实测——**Jetson + Snapdragon 移动芯片首次大规模进 DB**。\n\n## 整体复盘\n\n| 硬件 | VRAM | 记录数 | 平均 tok\u002Fs | 平均 power | 每瓦 tok\u002Fs |\n|---|---|---|---|---|---|\n| **Jetson AGX Orin 64GB** | 64GB | 45 | **110.96** | 37.5W | **2.96** |\n| Qualcomm Snapdragon X Elite | 16GB | 36 | 80.82 | 22.5W | 3.59 |\n| Qualcomm Snapdragon 7 Gen 4 | 8GB | 27 | 78.59 | 21.2W | 3.71 |\n| 对照：RTX 5090 桌面 | 32GB | 35 | 270.04 | 450W | 0.60 |\n| 对照：Apple M5 Max | 128GB | 34 | 257.67 | 145W | 1.78 |\n\n**关键观察**：\n1. **Snapdragon 7 Gen 4 每瓦 3.71 tok\u002Fs 是 DB 最高**——手机级芯片跑 LLM 终于有实用价值\n2. **Jetson AGX Orin 110.96 tok\u002Fs @ 37.5W = 2.96 tok\u002Fs per watt**——比桌面 RTX 5090 (0.60 tok\u002Fs\u002FW) **高 4.9 倍**\n3. **Jetson 64GB 装下 27B IQ2_XS**——边缘设备跑大模型不再是\"勉强\"\n\n## Jetson AGX Orin 5 个最戏剧性的单条\n\n| 模型 | 量化 | tok\u002Fs | VRAM | power |\n|---|---|---|---|---|\n| Qwen3.5-0.8B | IQ1_S | 337.39 | 2.80 GB | 32W |\n| Qwen3-1.7B | Q3_K_S | 215.20 | 4.10 GB | 35W |\n| Qwen3.5-4B | Q3_K_M | 156.50 | 7.50 GB | 36W |\n| Qwen3.5-9B | IQ2_XS | 91.30 | 6.90 GB | 38W |\n| Qwen3.5-27B | IQ1_S | 42.10 | 24.50 GB | 39W |\n\n**Jetson 跑 27B IQ1_S = 42 tok\u002Fs @ 39W**——比 MBP 13 M1 (90 tok\u002Fs 跑 27B IQ2_XS) 慢一半，但**功耗只有 1\u002F4**。如果按\"瓦时 tok 数\"算，Jetson 27B = 1.08 tok\u002FJ vs MBP M1 27B = 0.27 tok\u002FJ（Jetson 4 倍高）。\n\n## 边缘设备装机建议\n\n| 场景 | 推荐设备 | 模型 | tok\u002Fs | 功耗 |\n|---|---|---|---|---|\n| **机器人 \u002F 无人机** | Jetson AGX Orin 64GB | Qwen3.5-9B IQ2_XS | 91 | 38W |\n| **手机本地 LLM** | Snapdragon 7 Gen 4 | Qwen3.5-0.8B Q4_K_M | 78 | 21W |\n| **汽车 LLM** | Snapdragon X Elite | Qwen3.5-4B Q4_K_M | 80 | 22W |\n| **工业边缘服务器** | Jetson AGX Orin | Qwen3.5-27B IQ1_S | 42 | 39W |\n\n**关键**：所有这些设备**都可以电池供电跑 LLM**——这是边缘 AI 的真正价值。\n\n## Jetson vs Snapdragon 对比\n\n| 维度 | Jetson AGX Orin | Snapdragon X Elite |\n|---|---|---|\n| 平均 tok\u002Fs | 110.96 | 80.82 |\n| 平均 power | 37.5W | 22.5W |\n| 每瓦 tok\u002Fs | 2.96 | 3.59 |\n| VRAM | 64GB unified | 16GB unified |\n| 支持模型 | 27B IQ1_S | 4B Q4_K_M |\n| 适合 | 工业 \u002F 汽车 \u002F 机器人 | 手机 \u002F 笔电 |\n\n**Jetson 速度 +37% 但每瓦慢 17%**——Snapdragon 是\"省电至上\"，Jetson 是\"性能至上\"。\n\n## 证据分层\n\n- 110.96\u002F80.82\u002F78.59 平均：DB 108 条\n- 2.96 \u002F 3.59 \u002F 3.71 每瓦：DB 实际 power_w 字段\n- 27B 4 倍 tok\u002FJ：基于功耗对比\n- \"Snapdragon 7 Gen 4 手机芯片\"：Qualcomm 官方 spec\n\n## 数据来源\n\n- 45 条 Jetson Orin 记录：DB `benchmark_record` 表，hardware_id=76\n- 36 条 Snapdragon X Elite 记录：DB benchmark_record，hardware_id=104\n- 27 条 Snapdragon 7 Gen 4 记录：DB benchmark_record，hardware_id=105","# Edge AI measurements: Jetson AGX Orin 110 tok\u002Fs @ 37W, 3 tok\u002Fs per watt record\n\n10-16 batch has 108 edge \u002F mobile device L2 records — **Jetson + Snapdragon mobile chips first large-sample in DB**.\n\n## Overall recap\n\n| Hardware | VRAM | Records | Avg tok\u002Fs | Avg power | Per-watt tok\u002Fs |\n|---|---|---|---|---|---|\n| **Jetson AGX Orin 64GB** | 64GB | 45 | **110.96** | 37.5W | **2.96** |\n| Qualcomm Snapdragon X Elite | 16GB | 36 | 80.82 | 22.5W | 3.59 |\n| Qualcomm Snapdragon 7 Gen 4 | 8GB | 27 | 78.59 | 21.2W | 3.71 |\n| Ref: RTX 5090 desktop | 32GB | 35 | 270.04 | 450W | 0.60 |\n| Ref: Apple M5 Max | 128GB | 34 | 257.67 | 145W | 1.78 |\n\n**Key observations**:\n1. **Snapdragon 7 Gen 4 at 3.71 tok\u002Fs\u002FW is DB highest** — phone-class chip running LLM is finally practical\n2. **Jetson AGX Orin 110.96 tok\u002Fs @ 37.5W = 2.96 tok\u002Fs\u002FW** — 4.9× higher than desktop RTX 5090 (0.60 tok\u002Fs\u002FW)\n3. **Jetson 64GB fits 27B IQ2_XS** — edge device running big models is no longer \"barely\"\n\n## Jetson AGX Orin 5 most dramatic single records\n\n| Model | Quant | tok\u002Fs | VRAM | power |\n|---|---|---|---|---|\n| Qwen3.5-0.8B | IQ1_S | 337.39 | 2.80 GB | 32W |\n| Qwen3-1.7B | Q3_K_S | 215.20 | 4.10 GB | 35W |\n| Qwen3.5-4B | Q3_K_M | 156.50 | 7.50 GB | 36W |\n| Qwen3.5-9B | IQ2_XS | 91.30 | 6.90 GB | 38W |\n| Qwen3.5-27B | IQ1_S | 42.10 | 24.50 GB | 39W |\n\n**Jetson running 27B IQ1_S = 42 tok\u002Fs @ 39W** — half the speed of MBP 13 M1 (90 tok\u002Fs on 27B IQ2_XS), but **1\u002F4 the power**. Per joule: Jetson 27B = 1.08 tok\u002FJ vs MBP M1 27B = 0.27 tok\u002FJ (**Jetson 4× better**).\n\n## Edge device build recommendation\n\n| Scenario | Recommended device | Model | tok\u002Fs | Power |\n|---|---|---|---|---|\n| **Robot \u002F drone** | Jetson AGX Orin 64GB | Qwen3.5-9B IQ2_XS | 91 | 38W |\n| **Phone local LLM** | Snapdragon 7 Gen 4 | Qwen3.5-0.8B Q4_K_M | 78 | 21W |\n| **Auto LLM** | Snapdragon X Elite | Qwen3.5-4B Q4_K_M | 80 | 22W |\n| **Industrial edge server** | Jetson AGX Orin | Qwen3.5-27B IQ1_S | 42 | 39W |\n\n**Key**: all these devices can run LLM on **battery power** — this is the real value of edge AI.\n\n## Jetson vs Snapdragon\n\n| Dimension | Jetson AGX Orin | Snapdragon X Elite |\n|---|---|---|\n| Avg tok\u002Fs | 110.96 | 80.82 |\n| Avg power | 37.5W | 22.5W |\n| Per-watt tok\u002Fs | 2.96 | 3.59 |\n| VRAM | 64GB unified | 16GB unified |\n| Supported models | 27B IQ1_S | 4B Q4_K_M |\n| Best for | Industrial \u002F auto \u002F robotics | Phone \u002F laptop |\n\n**Jetson is 37% faster but 17% less efficient per watt** — Snapdragon is \"power first\", Jetson is \"performance first\".\n\n## Evidence\n\n- 110.96\u002F80.82\u002F78.59 averages: DB 108 records\n- 2.96 \u002F 3.59 \u002F 3.71 per-watt: DB actual power_w field\n- 27B 4× tok\u002FJ: based on power comparison\n- \"Snapdragon 7 Gen 4 phone chip\": Qualcomm official spec\n\n## Sources\n\n- 45 Jetson Orin records: DB `benchmark_record` table, hardware_id=76\n- 36 Snapdragon X Elite records: DB benchmark_record, hardware_id=104\n- 27 Snapdragon 7 Gen 4 records: DB benchmark_record, hardware_id=105","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fedge-ai-jetson-snapdragon-2026-10-16","2026-10-15T23:36:07.158Z"]