出典と計算方法
計算方法（推定）
- モデルサイズ = パラメータ数 × 量子化bit数 ÷ 8（Q4_K_M≈4.85bit, Q5_K_M≈5.69, Q8_0≈8.5, FP16=16）。gpt-ossはMXFP4（約4.25bit）ネイティブ配布のため量子化選択に関わらず固定
- KVキャッシュ = 2 × 層数 × KVヘッド数 × ヘッド次元 × 2byte × 文脈長（FP16）。Gemma 3・gpt-ossのスライディング窓層は窓長で頭打ち。構成値は各モデルのconfig.json
- 必要メモリ = モデルサイズ × 1.05 + KVキャッシュ + 1GB。Macは統合メモリの約72%をGPUで使える前提
- 生成速度 = メモリ帯域 × メーカー別効率 ÷ （アクティブパラメータ分のモデルサイズ）。効率は Llama 2 7B Q4_0 の公開実測の中央値: Apple 0.642 / Apple-hi 0.442 / NVIDIA 0.719 / AMD 0.767 / Intel 0.583（Apple-hiは帯域600GB/s超のMac）
- MoEモデルは理論値に届かないため、1トークンあたり+1.9msの固定オーバーヘッドを加算（#17621 のRTX 4090/5090実測から）
- 実測出典: #4167 (Apple) / #15013 (CUDA) / #10879 (Vulkan)
データ出典
- GeForce RTX 5090: https://www.techpowerup.com/gpu-specs/geforce-rtx-5090.c4216
- GeForce RTX 5080: https://www.techpowerup.com/gpu-specs/geforce-rtx-5080.c4217
- GeForce RTX 5070 Ti: https://www.techpowerup.com/gpu-specs/geforce-rtx-5070-ti.c4243
- GeForce RTX 5070: https://www.techpowerup.com/gpu-specs/geforce-rtx-5070.c4218
- GeForce RTX 5060 Ti 16 GB: https://www.techpowerup.com/gpu-specs/geforce-rtx-5060-ti-16-gb.c4292
- GeForce RTX 4090: https://www.techpowerup.com/gpu-specs/geforce-rtx-4090.c3889
- GeForce RTX 4080: https://www.techpowerup.com/gpu-specs/geforce-rtx-4080.c3888
- GeForce RTX 4070: https://www.techpowerup.com/gpu-specs/geforce-rtx-4070.c3924
- GeForce RTX 4060 Ti 16 GB: https://www.techpowerup.com/gpu-specs/geforce-rtx-4060-ti-16-gb.c4155
- GeForce RTX 4060: https://www.techpowerup.com/gpu-specs/geforce-rtx-4060.c4107
- GeForce RTX 3090: https://www.techpowerup.com/gpu-specs/geforce-rtx-3090.c3622
- GeForce RTX 3060 12 GB: https://www.techpowerup.com/gpu-specs/geforce-rtx-3060-12-gb.c3682
- Radeon RX 7900 XTX: https://www.techpowerup.com/gpu-specs/radeon-rx-7900-xtx.c3941
- Radeon RX 9070 XT: https://en.wikipedia.org/wiki/Radeon_RX_9000_series
- Arc B580: https://www.techpowerup.com/gpu-specs/arc-b580.c4244
- GeForce RTX 2060: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 2070 SUPER: https://en.wikipedia.org/wiki/GeForce_RTX_20_series
- GeForce RTX 2080 Ti: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 3060 Ti: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 3070: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 3080: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 3080 Ti: https://en.wikipedia.org/wiki/GeForce_RTX_30_series
- GeForce RTX 3090 Ti: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 4060 Ti 8 GB: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 4070 SUPER: https://en.wikipedia.org/wiki/GeForce_RTX_40_series
- GeForce RTX 4070 Ti: https://en.wikipedia.org/wiki/GeForce_RTX_40_series
- GeForce RTX 4070 Ti SUPER: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 4080 SUPER: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 5060: https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units
- GeForce RTX 4090 Laptop: https://en.wikipedia.org/wiki/GeForce_RTX_40_series
- GeForce RTX 5090 Laptop: https://en.wikipedia.org/wiki/GeForce_RTX_50_series
- Radeon RX 6700 XT: https://en.wikipedia.org/wiki/Radeon_RX_6000_series
- Radeon RX 6800: https://en.wikipedia.org/wiki/Radeon_RX_6000_series
- Radeon RX 6800 XT: https://en.wikipedia.org/wiki/Radeon_RX_6000_series
- Radeon RX 6900 XT: https://en.wikipedia.org/wiki/Radeon_RX_6000_series
- Radeon RX 6800M: https://en.wikipedia.org/wiki/Radeon_RX_6000_series
- Radeon RX 7600: https://en.wikipedia.org/wiki/Radeon_RX_7000_series
- Radeon RX 7700 XT: https://en.wikipedia.org/wiki/Radeon_RX_7000_series
- Radeon RX 7800 XT: https://en.wikipedia.org/wiki/Radeon_RX_7000_series
- Radeon RX 7900 GRE: https://en.wikipedia.org/wiki/Radeon_RX_7000_series
- Radeon RX 7900 XT: https://en.wikipedia.org/wiki/Radeon_RX_7000_series
- Radeon RX 9060 XT 16 GB: https://en.wikipedia.org/wiki/Radeon_RX_9000_series
- Radeon RX 9070: https://en.wikipedia.org/wiki/Radeon_RX_9000_series
- Radeon AI PRO R9700: https://www.amd.com/en/products/graphics/workstations/radeon-ai-pro/ai-9000-series/amd-radeon-ai-pro-r9700.html
- Arc A770 16 GB: https://en.wikipedia.org/wiki/Intel_Arc
- Arc A750: https://en.wikipedia.org/wiki/Intel_Arc
- Arc B570: https://en.wikipedia.org/wiki/Intel_Arc
- Apple M4 (base): https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/
- Apple M4 Pro: https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/
- Apple M4 Max: https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/
- Apple M3 Ultra (Mac Studio): https://www.apple.com/newsroom/2025/03/apple-unveils-new-mac-studio-the-most-powerful-mac-ever/
- Apple M2 Ultra: https://www.apple.com/newsroom/2023/06/apple-introduces-m2-ultra/
- Apple M1 Max: https://www.apple.com/newsroom/2021/10/introducing-m1-pro-and-m1-max-the-most-powerful-chips-apple-has-ever-built/
- Llama 3.1 8B Instruct: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
- Llama 3.3 70B Instruct: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
- Qwen2.5 7B Instruct: https://huggingface.co/Qwen/Qwen2.5-7B-Instruct
- Qwen2.5 14B Instruct: https://huggingface.co/Qwen/Qwen2.5-14B-Instruct
- Qwen2.5 32B Instruct: https://huggingface.co/Qwen/Qwen2.5-32B-Instruct
- Qwen2.5 72B Instruct: https://huggingface.co/Qwen/Qwen2.5-72B-Instruct
- Qwen3 30B-A3B: https://huggingface.co/Qwen/Qwen3-30B-A3B
- Qwen3 32B: https://huggingface.co/Qwen/Qwen3-32B
- Gemma 3 12B: https://huggingface.co/google/gemma-3-12b-it
- Gemma 3 27B: https://huggingface.co/google/gemma-3-27b-it
- Mistral Small 3.1 24B: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503
- Phi-4 14B: https://huggingface.co/microsoft/phi-4
- gpt-oss-20b: https://huggingface.co/openai/gpt-oss-20b
- gpt-oss-120b: https://huggingface.co/openai/gpt-oss-120b
- DeepSeek-R1-Distill-Qwen-14B: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
- Llama-3.1-Swallow-8B-Instruct: https://huggingface.co/tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.3
- Llama 2 7B (ベンチ基準): https://huggingface.co/meta-llama/Llama-2-7b-hf
広告・アフィリエイトについて
（準備中）当サイトはアフィリエイトプログラムに参加する予定です。リンク経由の購入で当サイトが紹介料を受け取る場合があります。