{"id":2256,"date":"2026-02-28T16:20:19","date_gmt":"2026-02-28T08:20:19","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2256"},"modified":"2026-02-28T16:20:19","modified_gmt":"2026-02-28T08:20:19","slug":"%e8%8b%b9%e6%9e%9cm4-neural-engine-38-tops%e5%be%88%e8%83%bd%e6%89%93%ef%bc%9f%e5%af%b9%e6%af%94%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97gpu%e4%ba%91%e4%b8%bb%e6%9c%ba%e5%ae%9e%e6%b5%8bllm%e6%8e%a8","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2256","title":{"rendered":"\u82f9\u679cM4 Neural Engine 38 TOPS\u5f88\u80fd\u6253\uff1f\u5bf9\u6bd4\u661f\u5b87\u667a\u7b97GPU\u4e91\u4e3b\u673a\u5b9e\u6d4bLLM\u63a8\u7406\u901f\u5ea6"},"content":{"rendered":"
\"\u82f9\u679cM4<\/figure>\n
\n

\u201cMac mini M4 \u65e5\u79df\u53ea\u8981 16.5 \u7f8e\u5143\uff0c\u672c\u5730\u8dd1 Llama-3-8B \u51b2\u4e0a\u70ed\u641c\uff01\u201d
\n\u8fc7\u53bb\u4e00\u5468\uff0c\u8fd9\u6761\u6d88\u606f\u50cf\u75c5\u6bd2\u4e00\u6837\u5728\u5f00\u53d1\u8005\u7fa4\u91cc\u6269\u6563\u3002\u82f9\u679c\u5b98\u65b9\u5ba3\u79f0 M4 \u7684 Neural Engine \u7b97\u529b 38 TOPS\uff0c\u542c\u4e0a\u53bb\u8db3\u591f\u628a\u5927\u6a21\u578b\u585e\u8fdb\u684c\u9762\u5c0f\u76d2\u5b50\u3002\u4e8e\u662f\uff0c\u4e00\u6279\u6279\u6781\u5ba2\u6d8c\u5411\u79df\u8d41\u5e73\u53f0\uff0c\u8bd5\u56fe\u7528\u4e00\u676f\u5496\u5561\u7684\u4ef7\u683c\u4f53\u9a8c\u201cApple AI\u201d\u3002\u53ef\u771f\u6b63\u8dd1\u8d77\u6765\u624d\u53d1\u73b0\uff1a\u9996 token \u5e73\u5747\u5ef6\u8fdf 1.8 \u79d2\uff0c\u5e76\u53d1 4 \u8def\u5c31\u6389\u5e27\uff0c\u98ce\u6247\u72c2\u8f6c\u529f\u8017\u98d9\u5230 39 W\uff0c\u6bcf\u5c0f\u65f6\u7535\u8d39\u518d\u52a0 0.06 \u7f8e\u5143\u2014\u2014\u5c1d\u9c9c\u6210\u672c\u6084\u6084\u7ffb\u500d\u3002<\/p>\n<\/blockquote>\n

\n

\u4e00\u300138 TOPS \u5230\u5e95\u4ec0\u4e48\u6c34\u5e73\uff1f<\/h2>\n

\u628a 38 TOPS \u6362\u7b97\u6210 GPU \u5708\u719f\u6089\u7684 FP16 \u7b97\u529b\uff0c\u22480.3\u00d7A100-80G\u3002
\nLlama-3-8B INT4 \u5b9e\u6d4b\u6570\u636e\u66f4\u76f4\u767d\uff1a<\/p>\n\n\n\n\n\n\n\n\n
\u6307\u6807<\/th>\nM4 16 GB<\/th>\nRTX 4090 GPU\u4e91\u4e3b\u673a<\/th>\n<\/tr>\n<\/thead>\n
\u9996 token \u5ef6\u8fdf<\/td>\n1.8 s<\/td>\n0.21 s<\/td>\n<\/tr>\n
\u5e76\u53d1 QPS<\/td>\n3.2<\/td>\n28<\/td>\n<\/tr>\n
\u5cf0\u503c\u529f\u8017<\/td>\n39 W<\/td>\n300 W\uff08\u4e91\u7aef\u7535\u8d39\u5e73\u53f0\u627f\u62c5\uff09<\/td>\n<\/tr>\n
\u65e5\u79df\u91d1<\/td>\n$16.5<\/td>\n$9.9\uff08\u661f\u5b87\u667a\u7b97\u6309\u91cf\uff09<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4e00\u53e5\u8bdd\uff1a\u7aef\u4fa7\u8f7b\u804a\u53ef\u4ee5\uff0c\u751f\u4ea7\u7ea7\u9ad8\u5e76\u53d1\u76f4\u63a5\u88ab\u4e91\u4e3b\u673a\u78be\u538b\u3002<\/p>\n

\n

\u4e8c\u3001\u4e3a\u4ec0\u4e48\u751f\u4ea7\u4ecd\u8981\u9009 GPU \u4e91\u4e3b\u673a\uff1f<\/h2>\n
    \n
  1. \n

    \u7b97\u529b\u5bc6\u5ea6<\/strong>
    \n \u5355\u5361 RTX 4090 82 TOPS\uff08\u5149\u8ffd\uff09+ 660 AI TOPS\uff08Tensor\uff09\uff0c\u662f M4 \u7684 17 \u500d\uff1bNVLink \u7248\u8fd8\u80fd\u6a2a\u5411\u5806\u53e0\uff0c\u8f7b\u677e\u628a 70 B \u6a21\u578b\u585e\u8fdb 24 GB \u663e\u5b58\u3002<\/p>\n<\/li>\n

  2. \n

    \u5f39\u6027\u5e76\u53d1<\/strong>
    \n \u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528 \u652f\u6301 1\uff5e8 \u5361\u5206\u949f\u7ea7\u4f38\u7f29\uff1a\u767d\u5929\u4f4e\u5cf0 1 \u5361\u505a\u63a8\u7406\uff0c\u591c\u95f4\u81ea\u52a8\u6269\u5bb9 8 \u5361\u8bad\u7ec3\uff0c\u6210\u672c\u8282\u7701 42%\u3002<\/p>\n<\/li>\n

  3. \n

    \u6570\u636e\u4e0e\u6a21\u578b\u201c\u79d2\u62c9\u201d<\/strong>
    \n \u672c\u5730 Mac \u9700\u5148\u4e0b\u8f7d 15 GB \u6743\u91cd\uff0c\u5343\u5146\u5bbd\u5e26\u4e5f\u8981 4 \u5206\u949f\uff1b\u661f\u5b87\u667a\u7b97\u5185\u7f6e\u300c\u5171\u4eab\u6a21\u578b\u4ed3\u5e93\u300d\uff0cLlama-3-8B\u3001Qwen-14B\u3001SDXL \u7b49 200+ \u516c\u5171\u6a21\u578b\u901a\u8fc7\u5185\u7f51\u6302\u8f7d\uff0c\u5f00\u673a\u5373\u53ef\u8c03\u7528\uff0c\u52a0\u8f7d\u901f\u5ea6 3 Gb\/s\uff0c\u96f6\u7b49\u5f85\u3001\u96f6\u672c\u5730\u5b58\u50a8\u3002<\/p>\n<\/li>\n

  4. \n

    \u65e0 GPU \u6a21\u5f0f\uff0c\u5f00\u53d1\u66f4\u7701<\/strong>
    \n \u5199\u4ee3\u7801\u3001\u8c03\u53c2\u6570\u9636\u6bb5\uff0c\u5f00\u4e00\u53f0 4 vCPU\u201c\u65e0 GPU\u201d\u5b9e\u4f8b\uff0c\u6bcf\u5206\u949f\u4ec5 0.02 \u5143\uff1b\u9700\u8981\u63a8\u7406\u65f6\u4e00\u952e\u70ed\u63d2\u5361\uff0c\u65e0\u9700\u91cd\u65b0\u90e8\u7f72\uff0c\u771f\u6b63\u505a\u5230\u201c\u8bad\u7ec3-\u63a8\u7406-\u8c03\u8bd5\u201d\u540c\u73af\u5883\u3002<\/p>\n<\/li>\n

  5. \n

    \u4ef7\u683c<\/strong>
    \n \u65b0\u7528\u6237\u6ce8\u518c\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0cRTX 4090 GPU\u4e91\u4e3b\u673a \u6309\u91cf\u6700\u4f4e 1.18 \u5143\/\u5c0f\u65f6\uff0c\u5305\u6708\u66f4\u4eab 6.8 \u6298\uff1b\u5bf9\u6bd4 Mac mini M4 \u65e5\u79df 120 \u5143\uff0c\u540c\u6837\u8dd1 8 \u5c0f\u65f6\u63a8\u7406\u4efb\u52a1\uff0c\u6210\u672c\u964d\u4f4e 60% \u4ee5\u4e0a\u3002<\/p>\n<\/li>\n<\/ol>\n

    \n

    \u4e09\u3001\u573a\u666f\u5316\u5efa\u8bae\uff1a\u7aef\u4fa7 vs \u4e91\u7aef<\/h2>\n\n\n\n\n\n\n\n\n
    \u573a\u666f<\/th>\n\u63a8\u8350\u65b9\u6848<\/th>\n<\/tr>\n<\/thead>\n
    \u4e2a\u4eba\u535a\u5ba2\u96c6\u6210 7 B \u95ee\u7b54<\/td>\nM4 \u672c\u5730\u5c1d\u9c9c\uff0c\u54cd\u5e94\u6162\u4e5f\u80fd\u63a5\u53d7<\/td>\n<\/tr>\n
    \u5c0f\u7a0b\u5e8f 1000 QPS \u5ba2\u670d<\/td>\n\u661f\u5b87\u667a\u7b97 4\u00d7RTX 4090 GPU\u4e91\u4e3b\u673a\uff0cAuto-Scaling \u81ea\u52a8\u5f39\u7f29<\/td>\n<\/tr>\n
    CV \u56e2\u961f\u6bcf\u5468\u8bad\u7ec3 300 G \u6570\u636e<\/td>\n\u661f\u5b87\u667a\u7b97 8\u00d7A100 \u5305\u6708\uff0c\u8d60\u9001 10 TB \u6301\u4e45\u5316\u4e91\u5b58\u50a8<\/td>\n<\/tr>\n
    \u9ad8\u6821\u8bfe\u7a0b 200 \u5b66\u751f\u5e76\u884c\u5b9e\u9a8c<\/td>\n\u6279\u91cf\u5f00\u901a\u300c\u65e0 GPU\u300d\u5b9e\u4f8b\uff0c\u968f\u8bfe\u5f00\u5173\uff0c\u4eba\u5747\u6210\u672c <3 \u5143\/\u8bfe\u65f6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
    \n

    \u56db\u3001\u5b9e\u6d4b\uff1a\u540c\u6837 8 \u5c0f\u65f6\u4e1a\u52a1\u9ad8\u5cf0<\/h2>\n