{"meta":{"name":"エッジAI実機検証DB（一次データ）","description":"自宅ラボの実機で同一手法により計測したエッジAIの一次データ。ローカルLLMの tok/s・日本語 char/s、NPU(Hailo)のビジョンFPS、Whisper音声認識のCER/実時間倍率、VLMの応答時間、公称vs実測、端末×タスクのFPSを収録。","version":"2026-07-03","dateModified":"2026-07-03","creator":{"name":"齊藤晃紀","url":"https://link-field.jp/","jobTitle":"GIS/AIエンジニア・測量士補","note":"南極観測隊参加、内閣府QZSS（みちびき）実績、QGIS公式プラグイン作者（6本）。","sameAs":["https://link-field.jp/","https://miru-lab.jp/","https://ai-edge-lab.com/"]},"publisher":"エッジAIラボ（ai-edge-lab.com）","canonicalUrl":"https://ai-edge-lab.com/edge-ai-verified/","jsonUrl":"https://ai-edge-lab.com/api/benchmarks/","csvUrl":"https://ai-edge-lab.com/api/benchmarks/?format=csv","reusePolicy":"本データは クリエイティブ・コモンズ 表示 4.0 国際（CC BY 4.0）で提供します。作者名と出典（本ページURL）を明記すれば、商用利用を含め自由に引用・転載・再配布できます。","licenseName":"CC BY 4.0","licenseUrl":"https://creativecommons.org/licenses/by/4.0/","measurementTechnique":"自宅ラボ実機での同一手法計測（ローカルLLM: Ollama /api/generate、音声/画像: transformers ASR・VLM fp16、NPUビジョン: hailortcli benchmark、NPU GenAI: hailo-ollama /api/chat）＋ブラウザ訪問端末の匿名集計。測定条件と限界を併記。","hardware":["RTX A6000 (48GB)","RTX 4060 Laptop","Mac mini M4","Jetson Orin Nano","Raspberry Pi 5 (+ AI HAT+ / Hailo-8L)","Raspberry Pi 5 (+ AI HAT+2 / Hailo-10H)","Coral USB Accelerator"]},"findings":[{"id":"gpu-myth","claim":"「GPUは10〜20倍速い」はエッジ/ノートでは出ない","metric":"最大 ×4.2（RTX 4060 / 7B）","detail":"同一機内で num_gpu=0 にしてCPU実行と比較すると、専用VRAMのRTX 4060ノートでも0.5Bで×2.8・7Bで×4.2が最大。統合メモリ機はさらに小さい。よく言う『10〜20倍』はエッジ級では観測されない（単発推論がメモリ帯域律速のため）。","tags":["LLM","ハード選定","実機"],"evidenceLabel":"“GPUは10〜20倍速い”は嘘？","evidenceHref":"/edge-ai-gpu-myth/"},{"id":"memory-decides","claim":"GPU優位の決め手は「メモリ構成」","metric":"統合メモリ ×1.1〜1.7／専用VRAM ×4超","detail":"Mac mini M4・Jetson Orin Nanoの統合メモリ(LPDDR5)はCPUと帯域を共有するためGPU優位が小さい（≦1〜1.7倍）。専用VRAM(GDDR6)を持つRTX 4060は効く。チップのTOPSより『メモリの種類と容量』が体感速度を決める。","tags":["LLM","ハード選定","実機"],"evidenceLabel":"実機ベンチ大全（GPU vs CPU）","evidenceHref":"/edge-ai-benchmarks/"},{"id":"vram-overflow","claim":"消費者GPUはVRAMを超えると逆に遅くなる","metric":"RTX 4060 / 14B で ×2.2 に低下","detail":"RTX 4060ノートは7Bで×4.2まで伸びるが、14BではVRAM容量を超えて×2.2まで落ちた。『載り切るか』が速度を左右する。大きいモデルは量子化やモデル縮小、またはVRAMの大きいGPUが要る。","tags":["LLM","ハード選定","実機"],"evidenceLabel":"実機ベンチ大全（モデルサイズ別）","evidenceHref":"/edge-ai-benchmarks/"},{"id":"npu-vision-only","claim":"汎用Ollama/llama.cpp（GGUF）はHailoのNPUを使わずCPU実行になる","metric":"YOLOv8m 76 FPS・ResNet-50 308 FPS（ビジョン側の真価）","detail":"Hailo-8/8L/10HはHailoRT/TAPPASという専用ランタイム向けにコンパイルしたモデルしか実行できず、汎用のOllama/llama.cppが読むGGUF形式はNPU非対応でCPU実行にフォールバックする（Pi 5のCPU実行は約11 tok/s）。ただしHailo-10H限定でHailo公式のGenAI専用ツール『hailo-ollama』を使えばNPU上でLLMを動かせる（→ npu-genai-hailo10h）。旧型のHailo-8Lにはこの経路がなく、ビジョン専用のまま。","tags":["ビジョン","実機","ハード選定"],"evidenceLabel":"Hailo-10Hレビュー","evidenceHref":"/hailo-review/"},{"id":"npu-genai-hailo10h","claim":"Hailo-10H(AI HAT+2)はHailo公式ツール経由でLLMも動く。ただしCPUより速いとは限らない","metric":"Llama3.2 1B: NPU 9.69 > CPU 8.72 tok/s／Qwen2.5 1.5B: NPU 7.23 < CPU 11.76 tok/s","detail":"Raspberry Pi 5 + AI HAT+2（Hailo-10H・HailoRT 5.3.0）に実機SSHで接続し、Hailo公式のOllama互換サーバ『hailo-ollama』(NPU)と同一機の素のOllama(CPU)を同一プロンプト・num_predict=80・ウォームアップ1回破棄後の単発計測で比較。Llama3.2 1BはNPUがCPUより約11%速いが、Qwen2.5 1.5BはCPUの方が約63%速く、「40 TOPSだから常に速い」は成立しない（小型LLMはメモリ帯域律速のため）。Qwen2.5のNPU出力にごく軽微な文字化けも確認。公式パッケージ(hailo-gen-ai-model-zoo)にVLMモデルは含まれず、Hailoが謳うVLM対応はこの範囲では未確認。単発計測のため複数回平均ではない点に注意。","tags":["LLM","実機","ハード選定"],"evidenceLabel":"Hailo-10Hレビュー（GenAI実測）","evidenceHref":"/hailo-review/","estimated":true},{"id":"spec-vs-real","claim":"カタログ値(TOPS/TGP)は実機と乖離する","metric":"TGP 公称115W → 実機 既定60W","detail":"同じRTX 4060 Laptopでも、薄型ノートの電力設計で実性能は変わる。公称TGP最大115Wに対し実機は最大95W・既定60W。最終判断は『公称TOPS』ではなく『タスク別の実測(tok/s・FPS・電力)』が確実。","tags":["ハード選定","実機"],"evidenceLabel":"公称 vs 実測（実機ベンチ大全 ③）","evidenceHref":"/edge-ai-benchmarks/"},{"id":"charsec","claim":"日本語LLMは tok/s でなく char/s で測れ","metric":"同52 tok/s で 英287／日81 char/s（約1/3.5）","detail":"日本語は1トークンが約1.5文字（英語は約5.5文字）。同じ52 tok/sでも日本語ユーザーが見る文字数は英語の約1/3.5。tok/s表記は日本語の実速度を約3.5倍過大に見せる。比較は char/s で。","tags":["LLM","日本語"],"evidenceLabel":"日本語LLMは char/s で測れ","evidenceHref":"/japanese-llm-charsec/"},{"id":"jp-purity","claim":"ローカルLLMの日本語に簡体字は混じらなかった","metric":"エッジ9モデル実測 = 簡体字 0","detail":"qwen2.5/qwen3/qwen3.5/gemma3・lfm（0.5〜8B）を簡体字混入検出器で実測。日本語指示あり・なし・thinking無効のいずれでも簡体字はゼロ（クリーン）。噂を鵜呑みにせず一次データで確認した。","tags":["LLM","日本語"],"evidenceLabel":"日本語に中国語は混じる？","evidenceHref":"/japanese-llm-purity/"},{"id":"webgpu-mobile-frag","claim":"WebGPUはモバイルで断片化。WASM落ち・発熱・クラッシュあり","metric":"訪問端末で匿名実測・集計中","detail":"対応外のGPU命令だとWASM(CPU)へフォールバックして遅く・発熱・タブ落ちが起こる。どの端末で動くかは多数の実機で集めるしかない。当サイトは訪問端末の結果（端末能力＋成否のみ・匿名）を集計し地図にしている。","tags":["ブラウザ"],"evidenceLabel":"WebGPU対応状況マトリクス","evidenceHref":"/webgpu-support-matrix/"},{"id":"pi5-readable","claim":"Pi 5の日本語LLMでも「黙読より速く」実用圏","metric":"約18 char/s（黙読 約7〜10字/秒）","detail":"Raspberry Pi 5(CPU)のqwen2.5:1.5Bは約18 char/s。日本語の黙読は概ね7〜10字/秒（経験則・要検証）なので、安価なエッジ機でも『読む速さより速い』＝待ち時間の少ない実用圏に入る。","tags":["実機","LLM","日本語"],"evidenceLabel":"実機ベンチ大全（日本語 char/s）","evidenceHref":"/edge-ai-benchmarks/","estimated":true},{"id":"mobile-vision","claim":"スマホ・タブレットでもビジョンAIは実用速度","metric":"物体検出 iPhone15Pro 47 / iPad Air M2 54 fps","detail":"ブラウザ(WebGPU/GPU)での物体検出・姿勢推定・画像分類は、モバイルでも概ね30〜60fps出る。『スマホでAIは無理』は誤解で、軽量モデルなら手元の端末で実用速度に達する。","tags":["ビジョン","ブラウザ"],"evidenceLabel":"端末ベンチ比較","evidenceHref":"/benchmarks/"},{"id":"npu-cheap-vision","claim":"安価なPiも「NPUを足す」とビジョンが実用化","metric":"RPi5 CPU 7fps → +Hailo 40fps（約5.7倍）","detail":"Raspberry Pi 5単体(CPU)の物体検出は約7fpsだが、AI HAT+(Hailo)を足すと約40fpsへ。1〜2万円台のエッジでも、NPUを併用すれば現場運用に足るFPSが出る。","tags":["ビジョン","実機","ハード選定"],"evidenceLabel":"実機ベンチ大全（NPU FPS）","evidenceHref":"/edge-ai-benchmarks/"},{"id":"browser-llm-gap","claim":"ブラウザLLMの速度はPCとスマホで大差","metric":"11〜57 tok/s（Pixel 8 11 ／ MacBook Pro 57）","detail":"同じブラウザLLMでも、ハイエンドPCは数十tok/sで快適、低価格スマホは一桁台で待ちが出る。LLM体験は端末性能の影響が大きい（日本語はtok/sでなくchar/sで体感を測ること）。","tags":["LLM","ブラウザ"],"evidenceLabel":"端末ベンチ比較","evidenceHref":"/benchmarks/"},{"id":"power-efficiency","claim":"電力あたりの性能はエッジ機が圧倒的に有利","metric":"実測FPS÷W：Pi+Hailo ~5.0 ／ Jetson ~3.8 ／ RTX4060 ~0.2","detail":"物体検出の実測FPSを消費電力で割ると、Pi5+Hailo(8W)やJetson(15W)はデスクトップGPU(300W)より桁違いに電力効率が高い。常時稼働・現場設置ではエッジ専用機が効く（公称W使用・経験則）。","tags":["ハード選定","実機"],"evidenceLabel":"実機ベンチ大全","evidenceHref":"/edge-ai-benchmarks/","estimated":true},{"id":"fl-quality-gate","claim":"連合学習はデータを出さず、毒データはゲートで棄却できる","metric":"精度を下げる更新は不採用（単調非減少ラチェット）","detail":"連合学習(FL)は端末のデータを送らずモデルだけを共有する。当ラボのFL実装はクライアント更新を品質ゲートで評価し、精度を下げる（悪意ある）更新を棄却。ブラウザで攻防を体験できる。","tags":["連合学習","ブラウザ","実機"],"evidenceLabel":"連合学習（FL）ガイド","evidenceHref":"/federated-learning/"},{"id":"quant-q4-fast-clean","claim":"量子化Q4は日本語品質を落とさずQ8より約1.5倍速い","metric":"qwen2.5:7b: Q4 165 ／ Q8 111 ／ fp16 65 char/s（簡体字いずれも0）","detail":"同一モデル(qwen2.5:7b)を量子化だけ変えてRTX A6000で実測。char/sはQ4_K_MがQ8_0の約1.49倍・fp16の約2.52倍。簡体字混入は3精度ともゼロ、かな比率0.67〜0.71で日本語として自然。文字レベルの日本語品質は量子化で崩れず、速度はQ4が有利（推論の正答率は別軸・要検証）。","tags":["LLM","日本語","実機"],"evidenceLabel":"日本語LLMは char/s で測れ（量子化別 実測）","evidenceHref":"/japanese-llm-charsec/","estimated":true},{"id":"whisper-realtime","claim":"Whisperは実時間の17〜35倍速で文字起こし（A6000）。turboが最速かつ最精度","metric":"27秒の日本語音声: turbo 34×(CER4.7%) ／ base 35×(5.5%) ／ small 17×(7.9%)","detail":"transformers の Whisper を RTX A6000(fp16) で実測。3モデルとも実時間の17〜35倍速（RTF 0.03〜0.06）で、large-v3-turboは最速クラスかつ最も正確。音声認識はLLM生成より軽く、エッジでも実用十分。※CERは27秒1クリップ(クリーンな合成音声)の参考値・要検証。","tags":["音声","実機","日本語"],"evidenceLabel":"実機ベンチ大全（Whisper 実測）","evidenceHref":"/edge-ai-benchmarks/","estimated":true},{"id":"cross-llm-jp","claim":"7-8B帯はqwen2.5≒llama3.1が日本語最速。簡体字混入は全ファミリーで0","metric":"char/s: qwen2.5:7b 169 ／ llama3.1:8b 162 ／ mistral:7b 138（簡体字いずれも0）","detail":"同一和文プロンプト・同サイズ帯(7-8B Q4)をRTX A6000で横断実測。char/sはqwen2.5≒llama3.1>mistral。簡体字混入はQwen系含め全モデルで0＝日本語純度は良好。一方で出力の癖は別で、mistralは英語混じり・llama3.1はMarkdown多用・thinking型(lfm2.5等)は思考が英語化し比較対象外。","tags":["LLM","日本語","実機"],"evidenceLabel":"実機ベンチ大全（ローカルLLM横断）","evidenceHref":"/edge-ai-benchmarks/","estimated":true},{"id":"vlm-image-understanding","claim":"VLMはA6000で画像を約1秒で正確に日本語説明。Qwen2.5-VL 3Bでも実用十分","metric":"応答時間: Qwen2.5-VL-3B 0.93s ／ 7B 1.61s（被写体を正確に同定）。SmolVLM-256Mは説明失敗","detail":"写真を「日本語で説明して」とVLMに依頼しRTX A6000で実測。Qwen2.5-VL 3B/7Bは約1〜1.6秒で被写体を正確に同定（猫2匹・テレビのリモコン・ソファ）、3Bが最速。極小のSmolVLM-256Mは速いが指示を反復するだけで説明できず＝VLMはモデルサイズが品質の下限を決める。","tags":["ビジョン","LLM","実機"],"evidenceLabel":"実機ベンチ大全（VLM 画像理解）","evidenceHref":"/edge-ai-benchmarks/","estimated":true}],"tables":{"llmGpuVsCpu":[{"dev":"ノートPC RTX 4060","mem":"専用VRAM(GDDR6)","size":"0.5B","gpu":"320","cpu":"115","x":"×2.8","note":""},{"dev":"ノートPC RTX 4060","mem":"専用VRAM(GDDR6)","size":"3B","gpu":"107","cpu":"27","x":"×4.0","note":""},{"dev":"ノートPC RTX 4060","mem":"専用VRAM(GDDR6)","size":"7B","gpu":"52","cpu":"13","x":"×4.2","note":"最大"},{"dev":"ノートPC RTX 4060","mem":"専用VRAM(GDDR6)","size":"14B","gpu":"13.9","cpu":"6.4","x":"×2.2","note":"VRAM超過で低下"},{"dev":"Mac mini M4","mem":"統合(LPDDR5)","size":"0.5B","gpu":"157","cpu":"138","x":"×1.1","note":""},{"dev":"Mac mini M4","mem":"統合(LPDDR5)","size":"7B","gpu":"22.4","cpu":"15.6","x":"×1.4","note":""},{"dev":"Mac mini M4","mem":"統合(LPDDR5)","size":"14B","gpu":"11.7","cpu":"7.1","x":"×1.7","note":""},{"dev":"Jetson Orin Nano","mem":"統合(LPDDR5)","size":"0.8B","gpu":"9.2","cpu":"9.9","x":"×0.9","note":"GPU≒CPU"},{"dev":"Jetson Orin Nano","mem":"統合(LPDDR5)","size":"2B","gpu":"7.0","cpu":"11.3","x":"×0.6","note":"GPUが遅い"},{"dev":"Raspberry Pi 5","mem":"GPUなし(CPU)","size":"1.5B","gpu":"—","cpu":"11.4","x":"—","note":"CPUのみ"}],"npuVisionFps":[{"model":"YOLOv11m","task":"物体検出","fps":"71"},{"model":"YOLOv8m","task":"物体検出","fps":"76"},{"model":"YOLOv8s-pose","task":"姿勢推定","fps":"157"},{"model":"ResNet-50","task":"画像分類","fps":"308"}],"japaneseCharSec":[{"dev":"ノートPC RTX 4060 (GPU)","model":"qwen2.5:0.5B","toks":"325","chars":"520","cpt":"1.6"},{"dev":"ノートPC RTX 4060 (GPU)","model":"qwen2.5:7B","toks":"52","chars":"81","cpt":"1.55"},{"dev":"Mac mini M4 (Metal)","model":"qwen2.5:0.5B","toks":"154","chars":"234","cpt":"1.51"},{"dev":"Mac mini M4 (Metal)","model":"qwen2.5:7B","toks":"22.5","chars":"34","cpt":"1.50"},{"dev":"Jetson Orin Nano","model":"qwen2.5:1.5B","toks":"35","chars":"56","cpt":"1.58"},{"dev":"Raspberry Pi 5 (CPU)","model":"qwen2.5:1.5B","toks":"11","chars":"18","cpt":"1.60"}],"specVsReal":[{"item":"TGP（電力上限）","spec":"最大115W（35〜115W可変）","real":"最大95W／既定60W"},{"item":"GPUブーストクロック","spec":"最大 2370MHz","real":"負荷時 約2600MHz（公称超）"},{"item":"消費電力","spec":"—","real":"アイドル12.86W → LLM生成中ピーク約70W"},{"item":"AI性能","spec":"233 AI TOPS（INT8理論ピーク）","real":"LLM 0.5B ≈ 320 tok/s（GPU）"}],"whisperAsr":[{"model":"whisper-base","proc":"0.79s","x":"34.6×","rtf":"0.029","cer":"5.5"},{"model":"whisper-small","proc":"1.57s","x":"17.3×","rtf":"0.058","cer":"7.9"},{"model":"whisper-large-v3-turbo","proc":"0.79s","x":"34.4×","rtf":"0.029","cer":"4.7"}],"crossLlmJapanese":[{"model":"qwen2.5:7b","origin":"Alibaba","chars":"169","toks":"119","kana":"0.63","simp":"0","note":""},{"model":"llama3.1:8b","origin":"Meta","chars":"162","toks":"113","kana":"0.64","simp":"0","note":"Markdown多用"},{"model":"mistral:7b","origin":"Mistral","chars":"138","toks":"124","kana":"0.74","simp":"0","note":"英語混じり"}],"vlmImage":[{"model":"Qwen2.5-VL-3B","sec":"0.93s","out":"33字","ok":"✅ 正確（猫2匹・リモコン・ソファを同定）"},{"model":"Qwen2.5-VL-7B","sec":"1.61s","out":"64字","ok":"✅ 正確・より詳細"},{"model":"SmolVLM-256M","sec":"1.74s","out":"30字","ok":"❌ 指示を反復し説明できず（256Mは小さすぎ）"}]},"deviceMatrix":[{"deviceId":"macbook-air-m3","device":"MacBook Air M3","chip":"Apple M3 (18 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":164800,"powerWatts":30,"task":"object-detection","taskLabel":"物体検出","fps":57,"tokensPerSec":null,"latencyMs":17,"backend":"gpu","estimated":false},{"deviceId":"macbook-air-m3","device":"MacBook Air M3","chip":"Apple M3 (18 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":164800,"powerWatts":30,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":54,"tokensPerSec":null,"latencyMs":18,"backend":"gpu","estimated":false},{"deviceId":"macbook-air-m3","device":"MacBook Air M3","chip":"Apple M3 (18 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":164800,"powerWatts":30,"task":"image-classification","taskLabel":"画像分類","fps":60,"tokensPerSec":null,"latencyMs":8,"backend":"gpu","estimated":false},{"deviceId":"macbook-air-m3","device":"MacBook Air M3","chip":"Apple M3 (18 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":164800,"powerWatts":30,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":35,"latencyMs":null,"backend":"webgpu","estimated":false},{"deviceId":"macbook-pro-m4pro","device":"MacBook Pro M4 Pro","chip":"Apple M4 Pro (38 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":298800,"powerWatts":70,"task":"object-detection","taskLabel":"物体検出","fps":60,"tokensPerSec":null,"latencyMs":12,"backend":"gpu","estimated":false},{"deviceId":"macbook-pro-m4pro","device":"MacBook Pro M4 Pro","chip":"Apple M4 Pro (38 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":298800,"powerWatts":70,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":60,"tokensPerSec":null,"latencyMs":14,"backend":"gpu","estimated":false},{"deviceId":"macbook-pro-m4pro","device":"MacBook Pro M4 Pro","chip":"Apple M4 Pro (38 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":298800,"powerWatts":70,"task":"image-classification","taskLabel":"画像分類","fps":60,"tokensPerSec":null,"latencyMs":5,"backend":"gpu","estimated":false},{"deviceId":"macbook-pro-m4pro","device":"MacBook Pro M4 Pro","chip":"Apple M4 Pro (38 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":298800,"powerWatts":70,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":57,"latencyMs":null,"backend":"webgpu","estimated":false},{"deviceId":"windows-rtx4060","device":"Windows PC (RTX 4060)","chip":"RTX 4060 (232 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":150000,"powerWatts":300,"task":"object-detection","taskLabel":"物体検出","fps":60,"tokensPerSec":null,"latencyMs":10,"backend":"webgpu","estimated":false},{"deviceId":"windows-rtx4060","device":"Windows PC (RTX 4060)","chip":"RTX 4060 (232 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":150000,"powerWatts":300,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":60,"tokensPerSec":null,"latencyMs":11,"backend":"webgpu","estimated":false},{"deviceId":"windows-rtx4060","device":"Windows PC (RTX 4060)","chip":"RTX 4060 (232 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":150000,"powerWatts":300,"task":"image-classification","taskLabel":"画像分類","fps":60,"tokensPerSec":null,"latencyMs":4,"backend":"webgpu","estimated":false},{"deviceId":"windows-rtx4060","device":"Windows PC (RTX 4060)","chip":"RTX 4060 (232 TOPS)","category":"PC / ノートPC","isOwned":false,"priceJPY":150000,"powerWatts":300,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":52,"latencyMs":null,"backend":"webgpu","estimated":false},{"deviceId":"iphone-15-pro","device":"iPhone 15 Pro","chip":"A17 Pro (35 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":159800,"powerWatts":5,"task":"object-detection","taskLabel":"物体検出","fps":47,"tokensPerSec":null,"latencyMs":22,"backend":"gpu","estimated":false},{"deviceId":"iphone-15-pro","device":"iPhone 15 Pro","chip":"A17 Pro (35 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":159800,"powerWatts":5,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":42,"tokensPerSec":null,"latencyMs":25,"backend":"gpu","estimated":false},{"deviceId":"iphone-15-pro","device":"iPhone 15 Pro","chip":"A17 Pro (35 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":159800,"powerWatts":5,"task":"image-classification","taskLabel":"画像分類","fps":57,"tokensPerSec":null,"latencyMs":10,"backend":"gpu","estimated":false},{"deviceId":"iphone-15-pro","device":"iPhone 15 Pro","chip":"A17 Pro (35 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":159800,"powerWatts":5,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":20,"latencyMs":null,"backend":"webgpu","estimated":false},{"deviceId":"galaxy-s24","device":"Galaxy S24","chip":"Snapdragon 8 Gen 3 (45 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":124700,"powerWatts":5,"task":"object-detection","taskLabel":"物体検出","fps":42,"tokensPerSec":null,"latencyMs":25,"backend":"gpu","estimated":false},{"deviceId":"galaxy-s24","device":"Galaxy S24","chip":"Snapdragon 8 Gen 3 (45 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":124700,"powerWatts":5,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":37,"tokensPerSec":null,"latencyMs":28,"backend":"gpu","estimated":false},{"deviceId":"galaxy-s24","device":"Galaxy S24","chip":"Snapdragon 8 Gen 3 (45 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":124700,"powerWatts":5,"task":"image-classification","taskLabel":"画像分類","fps":54,"tokensPerSec":null,"latencyMs":12,"backend":"gpu","estimated":false},{"deviceId":"galaxy-s24","device":"Galaxy S24","chip":"Snapdragon 8 Gen 3 (45 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":124700,"powerWatts":5,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":16,"latencyMs":null,"backend":"webgpu","estimated":false},{"deviceId":"pixel-8","device":"Pixel 8","chip":"Tensor G3 (10 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":82280,"powerWatts":5,"task":"object-detection","taskLabel":"物体検出","fps":30,"tokensPerSec":null,"latencyMs":35,"backend":"gpu","estimated":false},{"deviceId":"pixel-8","device":"Pixel 8","chip":"Tensor G3 (10 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":82280,"powerWatts":5,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":26,"tokensPerSec":null,"latencyMs":40,"backend":"gpu","estimated":false},{"deviceId":"pixel-8","device":"Pixel 8","chip":"Tensor G3 (10 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":82280,"powerWatts":5,"task":"image-classification","taskLabel":"画像分類","fps":45,"tokensPerSec":null,"latencyMs":18,"backend":"gpu","estimated":false},{"deviceId":"pixel-8","device":"Pixel 8","chip":"Tensor G3 (10 TOPS)","category":"スマートフォン","isOwned":false,"priceJPY":82280,"powerWatts":5,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":11,"latencyMs":null,"backend":"webgpu","estimated":false},{"deviceId":"ipad-air-m2","device":"iPad Air M2","chip":"Apple M2 (15 TOPS)","category":"タブレット","isOwned":false,"priceJPY":98800,"powerWatts":10,"task":"object-detection","taskLabel":"物体検出","fps":54,"tokensPerSec":null,"latencyMs":18,"backend":"gpu","estimated":false},{"deviceId":"ipad-air-m2","device":"iPad Air M2","chip":"Apple M2 (15 TOPS)","category":"タブレット","isOwned":false,"priceJPY":98800,"powerWatts":10,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":50,"tokensPerSec":null,"latencyMs":20,"backend":"gpu","estimated":false},{"deviceId":"ipad-air-m2","device":"iPad Air M2","chip":"Apple M2 (15 TOPS)","category":"タブレット","isOwned":false,"priceJPY":98800,"powerWatts":10,"task":"image-classification","taskLabel":"画像分類","fps":59,"tokensPerSec":null,"latencyMs":9,"backend":"gpu","estimated":false},{"deviceId":"ipad-air-m2","device":"iPad Air M2","chip":"Apple M2 (15 TOPS)","category":"タブレット","isOwned":false,"priceJPY":98800,"powerWatts":10,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":30,"latencyMs":null,"backend":"webgpu","estimated":false},{"deviceId":"jetson-orin-nano-super","device":"Jetson Orin Nano Super","chip":"Ampere GPU (67 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":75000,"powerWatts":15,"task":"object-detection","taskLabel":"物体検出","fps":57,"tokensPerSec":null,"latencyMs":15,"backend":"gpu","estimated":false},{"deviceId":"jetson-orin-nano-super","device":"Jetson Orin Nano Super","chip":"Ampere GPU (67 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":75000,"powerWatts":15,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":55,"tokensPerSec":null,"latencyMs":18,"backend":"gpu","estimated":false},{"deviceId":"jetson-orin-nano-super","device":"Jetson Orin Nano Super","chip":"Ampere GPU (67 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":75000,"powerWatts":15,"task":"image-classification","taskLabel":"画像分類","fps":60,"tokensPerSec":null,"latencyMs":8,"backend":"gpu","estimated":false},{"deviceId":"jetson-orin-nano-super","device":"Jetson Orin Nano Super","chip":"Ampere GPU (67 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":75000,"powerWatts":15,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":27,"latencyMs":null,"backend":"gpu","estimated":false},{"deviceId":"rpi5-ai-hat","device":"RPi 5 + AI HAT+","chip":"Hailo-8L (13 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":23000,"powerWatts":8,"task":"object-detection","taskLabel":"物体検出","fps":40,"tokensPerSec":null,"latencyMs":25,"backend":"npu","estimated":false},{"deviceId":"rpi5-ai-hat","device":"RPi 5 + AI HAT+","chip":"Hailo-8L (13 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":23000,"powerWatts":8,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":32,"tokensPerSec":null,"latencyMs":30,"backend":"npu","estimated":false},{"deviceId":"rpi5-ai-hat","device":"RPi 5 + AI HAT+","chip":"Hailo-8L (13 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":23000,"powerWatts":8,"task":"image-classification","taskLabel":"画像分類","fps":47,"tokensPerSec":null,"latencyMs":15,"backend":"npu","estimated":false},{"deviceId":"rpi5-ai-hat","device":"RPi 5 + AI HAT+","chip":"Hailo-8L (13 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":23000,"powerWatts":8,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":null,"latencyMs":null,"backend":"npu","estimated":false},{"deviceId":"rpi5-cpu","device":"Raspberry Pi 5","chip":"BCM2712 (CPU only)","category":"エッジデバイス","isOwned":true,"priceJPY":13000,"powerWatts":5,"task":"object-detection","taskLabel":"物体検出","fps":7,"tokensPerSec":null,"latencyMs":120,"backend":"cpu","estimated":false},{"deviceId":"rpi5-cpu","device":"Raspberry Pi 5","chip":"BCM2712 (CPU only)","category":"エッジデバイス","isOwned":true,"priceJPY":13000,"powerWatts":5,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":5,"tokensPerSec":null,"latencyMs":180,"backend":"cpu","estimated":false},{"deviceId":"rpi5-cpu","device":"Raspberry Pi 5","chip":"BCM2712 (CPU only)","category":"エッジデバイス","isOwned":true,"priceJPY":13000,"powerWatts":5,"task":"image-classification","taskLabel":"画像分類","fps":12,"tokensPerSec":null,"latencyMs":80,"backend":"cpu","estimated":false},{"deviceId":"rpi5-cpu","device":"Raspberry Pi 5","chip":"BCM2712 (CPU only)","category":"エッジデバイス","isOwned":true,"priceJPY":13000,"powerWatts":5,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":2,"latencyMs":null,"backend":"cpu","estimated":false},{"deviceId":"coral-usb","device":"Coral USB Accelerator","chip":"Edge TPU (4 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":10000,"powerWatts":2,"task":"object-detection","taskLabel":"物体検出","fps":25,"tokensPerSec":null,"latencyMs":40,"backend":"npu","estimated":true},{"deviceId":"coral-usb","device":"Coral USB Accelerator","chip":"Edge TPU (4 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":10000,"powerWatts":2,"task":"pose-estimation","taskLabel":"ポーズ推定","fps":20,"tokensPerSec":null,"latencyMs":50,"backend":"npu","estimated":true},{"deviceId":"coral-usb","device":"Coral USB Accelerator","chip":"Edge TPU (4 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":10000,"powerWatts":2,"task":"image-classification","taskLabel":"画像分類","fps":35,"tokensPerSec":null,"latencyMs":28,"backend":"npu","estimated":true},{"deviceId":"coral-usb","device":"Coral USB Accelerator","chip":"Edge TPU (4 TOPS)","category":"エッジデバイス","isOwned":true,"priceJPY":10000,"powerWatts":2,"task":"chat-ai","taskLabel":"チャットAI","fps":null,"tokensPerSec":null,"latencyMs":null,"backend":"npu","estimated":true}]}